Velocity Ascent

Looking toward tomorrow today

Joe Skopek

The Plumbing Nobody’s Building: Agentic Web’s Missing Infrastructure

· ·

When your agents can’t prove they’re trustworthy, your business can’t either.

Everyone is racing to build smarter AI agents. Fewer people are asking the harder question: once we have trillions of them, how do they find each other, verify each other’s credentials, and collaborate without tearing the whole system apart?

Sandeep Mahendru – an Engineering Leader and Senior System Architect at Google, where he builds AI-driven solutions for the financial market – recently published a piece on Medium that lays out the clearest architecture I’ve seen for answering that question. It’s called Proposal for the Global Backbone of the Agentic Web, and it’s worth a careful read.

“We are currently building millions of specialized agents, yet we have no global address book to find them, no shared language to verify their skills, and no scalable security layer to ensure they aren’t actively sabotaging our workflows.”

I am walking through it in this post as it maps to work we’ve been doing at Velocity Ascent – and because the ideas in it have real consequences for anyone thinking seriously about where enterprise AI is heading.

The “Wild West” Moment

Sandeep opens with a frame I find hard to argue with:

“We are currently building millions of specialized agents, yet we have no global address book to find them, no shared language to verify their skills, and no scalable security layer to ensure they aren’t actively sabotaging our workflows.”

This is the gap. It’s not a technical gap in the narrow sense – the individual pieces exist. Large language models are remarkable. Agent frameworks are maturing. But the connective tissue between agents – the infrastructure that makes a network of agents trustworthy and useful rather than just chaotic – doesn’t yet exist at scale.

Sandeep calls this the “Cisco Moment” of the Agentic Web. The analogy is deliberate: the winners in the next era of AI won’t be defined by who builds the smartest model. They’ll be defined by who builds the backbone; the discovery, routing, and trust layer – that lets trillions of specialized agents find and work with each other safely.

* A “Cisco moment” is a stock market analogy referring to the dot-com era when networking giant Cisco Systems saw its stock plummet roughly 80% after telecom companies stopped buying its hardware. Today, financial analysts use the term to warn that massive capital expenditures in AI might create a similar bubble, particularly drawing parallels between Cisco in 2000 and Nvidia today. [1, 2, 3, 4]

From Code-First to Intent-First

The shift Sandeep describes isn’t just infrastructural – it’s epistemological.

Traditional software development is code-first. A developer translates a goal into explicit instructions, one procedure at a time. What’s emerging is intent-first orchestration: you describe what you want, and an underlying intelligent infrastructure figures out which combination of specialized agents can execute it.

“Instead of building monolithic software silos, developers are now focused on defining high-level goals that are executed by a ‘quilt’ of autonomous, specialized agents.”

The “quilt” metaphor is worth sitting with. Not a single God Model that does everything, but a federated patchwork of best-in-class experts – each sovereign, each specialized, each discoverable. The question is how you stitch them together in real time.

That stitching mechanism is what Sandeep calls the Brain.

The Brain: Semantic Intent Parsing at Scale

The Brain is the orchestration intelligence at the center of this architecture. Think of it less like a search engine and more like a very sophisticated dispatcher.

Traditional keyword search is too blunt for this. If a user’s request is “help me optimize our supply chain for Q4,” a keyword search returns documents. An intelligent Brain decomposes that intent into a structured, multi-agent workflow – logistics expert, finance agent, demand forecasting agent – and routes each sub-task to the best available specialist in real time.

Image: Sandeep Mahendru on Medium.

Critically, Sandeep argues against a centralized bottleneck architecture:

“A standardized metadata routing protocol allows secure, local orchestrators to decompose human requests into multi-agent workflows while maintaining strict context isolation.”

This is a subtle but important design principle. Discovery doesn’t have to be centralized to be powerful. A federated model, where a lightweight metadata layer indexes agent capabilities across independently governed registries, can be both scalable and resilient.

The Discovery Gap: Why Agent Cards Aren’t Enough

The current standard for agent interoperability is A2A (Agent-to-Agent protocol), which gives every agent a structured “Agent Card” – essentially a business card declaring its capabilities, authentication requirements, and communication standards. A2A is built on established web infrastructure: JSON-RPC 2.0 over HTTP/S, OAuth 2.0, mTLS. It’s solid, enterprise-grade, and open.

But Agent Cards are self-declared. And self-declaration, as anyone who’s reviewed a vendor pitch deck knows, is not the same as verified performance.


AgentFacts: Report Cards, Not Business Cards

“Trust is no longer a ‘claim’ or a static audit; it’s an audited metric signed by third-party authorities and constantly tested via automated prompting and local consensus.”

AgentFacts are the proposed evolution beyond Agent Cards. Where an Agent Card says “I am a financial analysis agent,” an AgentFact is a cryptographically signed, W3C Verifiable Credential that says “this agent has been independently audited, has completed 12,000 successful tasks with a 97.3% accuracy rate, and has zero confirmed hallucinations in regulated output contexts.”

The distinction matters enormously in enterprise settings. When an agent is making decisions that touch compliance, financial transactions, or patient care, “trust me, I’m good at this” is not an acceptable credential.

Byzantine Fault Tolerance: In-Flight Security

Even cryptographically signed AgentFacts don’t solve the full security problem. Sandeep’s architecture adds a layer that I think most people building in this space haven’t adequately grappled with: Byzantine Fault Tolerance.

The analogy in the piece is the best summary I can offer:

“A Secure Federated Registry is like a ‘No Fly List’ – it keeps known bad actors out. But Byzantine Fault Tolerance is like ‘In-Flight Security’ – it protects the passengers even if someone with a valid ticket turns out to be a threat.”

BFT, in this context, means triangulating outputs across multiple independent agents and using quorum voting to detect when a “trusted” actor is hallucinating, has been compromised, or is simply wrong. Because traditional consensus mechanisms add latency, Sandeep proposes economic staking and slashing as the enforcement layer – agents put something at risk when they participate, and bad actors lose it.

This is crypto-native thinking applied to AI infrastructure. Whether or not you have a background in distributed systems, the principle is intuitive: skin in the game changes behavior.

Cognitive Liquidity: The AI-NPM Model

The long-term economic vision Sandeep lays out is what he calls Cognitive Liquidity – the ability to “import” verified intelligence as easily as a developer imports a software library today.

“By integrating discovered agents directly into low-code environments… developers can ‘import’ verified intelligence as easily as they import software libraries today.”

This is the AI-NPM model. Just as npm transformed software development by making reusable, versioned, community-maintained code packages trivially accessible, a mature Agentic Web would make reusable, verifiable, audited AI capabilities trivially accessible – streamable on demand, licensed transparently, and composable into complex workflows without custom integration work.

The infrastructure question isn’t just technical. It’s economic. Whoever builds the registry, the Brain, and the trust layer controls the tollbooth for an enormous amount of future AI commerce.

* In the context of AI, NPM (Node Package Manager) primarily refers to its role as the distribution hub for AI developer tools, libraries, and autonomous agents. While originally built for JavaScript, it has evolved into a “universal distribution platform” for the AI ecosystem.

Open Standards and the Public Good Mandate

One of the most important arguments in the piece is the one Sandeep makes almost as a footnote: this infrastructure needs to be governed by neutral bodies, not owned by any single platform.

A2A is already moving in this direction through the Linux Foundation and related open standard initiatives. NANDA – the Naming and Discovery Architecture – is one of the projects building the federated registry layer. If any single company owns the “Global Address Book,” the network effect becomes a chokehold. Open governance is the only model that scales without creating the kind of rent-extraction that would strangle the ecosystem before it matures.

A Note on Where This Work Lives

Sandeep Mahendru and I are both members of the NYC Chapter of the NANDA Project, an initiative out of MIT Media Lab focused on building the open naming and discovery layer for the Agentic Web. The work Sandeep describes isn’t purely theoretical – it’s the kind of architecture we’re actively thinking about in that community, and it’s influencing real infrastructure decisions being made right now.

The Agentic Web isn’t coming. It’s being built. The question is whether the backbone gets built right.

Why Agent Trust Matters to the C-Suite

If you’re a technology leader or founder thinking about AI strategy, Sandeep’s framework suggests a few reorienting questions:

  • Are you building agents that will be discoverable by other agents? Or are you building another closed silo?
  • How are you establishing verifiable trust in your AI outputs – not just internally, but in ways that external systems could validate?
  • Are your AI investments positioned for the intent-first world, or optimized for the code-first world you’re leaving behind?

The companies that get this right early will have an enormous structural advantage. The backbone of the Agentic Web is being laid now, and the patterns being established will be hard to reverse.

Sandeep Mahendru’s original article, Proposal for the Global Backbone of the Agentic Web, is published on Medium. Sandeep is an Engineering Leader and Senior System Architect at Google, where he builds AI-driven innovative solutions for financial markets.

Joe Skopek is the founder of Velocity Ascent, an AI-first innovation consultancy based in New York. Both are part of the Leadership Team of the NYC Chapter of the NANDA Project from MIT Media Lab.

How Do You Scope Work You’re Not Allowed to See? You Build Agents.

· ·

Analyze Everything. Read Nothing. A court-ordered constraint became the design brief for a new class of agentic pipeline – built for legal, compliance, and regulated document work at any scale.

The most interesting systems get built under impossible constraints.

In early 2026, Velocity Ascent was engaged to support a high-volume foreign language legal document translation project under active litigation. A New York City translation company had been retained by an international law firm operating under a court-issued protective order. The requirement was precise and non-negotiable: produce an accurate translation scope and cost estimate across thousands of pages of scanned source documents – without any member of the project team reading the underlying content.

Velocity Ascent designs bespoke agentic systems that maximize what AI can do autonomously – while staying precisely within the legal, regulatory, and compliance requirements specific to each client’s situation.

The documents existed. The estimate had to be produced. The protective order governed exactly what could and could not be touched. The gap between those two facts is where the engineering began.

Velocity Ascent designs bespoke agentic systems that maximize what AI can do autonomously – while staying precisely within the legal, regulatory, and compliance requirements specific to each client’s situation. The system described here was built for one engagement. The architecture behind it is built for any.

What emerged from that constraint is an agentic pipeline we believe has implications well beyond a single translation project – for any organization that needs to analyze, classify, or route sensitive document corpora without exposing their contents to human reviewers.”

The Compliance Problem Nobody Talks About: What happens when the requirement is to analyze without access?

Most regulated document workflows assume that the people building the pipeline can see what’s inside it. Translation firms scope work by reading samples. Legal teams estimate review hours by examining files. Compliance officers assess risk by opening documents.

Batch Analyzer – Ingests a compressed document batch, runs OCR where needed, classifies each file by type and complexity, and outputs a structured workbook ready for quoting and assignment.

Court orders, privilege designations, data sovereignty rules, and cross-border regulatory requirements increasingly break that assumption. The analyst cannot read the document. The estimator cannot open the file. But the work still has to be scoped, quoted, and delivered accurately.

Traditional approaches fail here in a specific way: they treat “review the documents” and “estimate the scope” as a single inseparable step. If you cannot do the first, you cannot do the second. The project stalls, costs inflate, or the constraint gets quietly worked around in ways that create downstream exposure.

Batch Extractor – Takes a client order specifying document IDs, maps each ID to a physical page within the source files, flags any unresolvable references for review, and extracts only the targeted pages into organized output folders.

Agentic pipelines solve this by separating observation from comprehension. A properly designed agent can characterize a document – page count, word density, language composition, structural complexity, document type – without surfacing a single line of content to any human reviewer. The observation layer and the content layer never meet.

Agentic Architecture: The Double Garden Wall Applied to Document Intelligence

The Double Garden Wall is an architecture we first developed for our ethical AI image generation tool – a system that needed to guarantee CC0 provenance on every training asset without relying on human memory or manual spot-checks. The principle transfers directly to regulated document handling.

Double Garden Wall architecture: specialized agents operating within layered compliance boundaries, where every document is characterized structurally and no content crosses to any human reviewer.

The outer wall defines what enters the system at all: only authorized document batches, governed by a court reference number logged at intake. Nothing flows in without an audit trail attached to it. The inner wall ensures that what exists inside the system – the actual document content – never crosses to the analysis agents. Statistical signals are sufficient. Page counts, word densities, language composition, Bates-to-page mappings, and structural classification are all derivable without any agent reading the underlying text.

The architecture employs two core agents working in sequence:

The Batch Analyzer ingests a ZIP archive of scanned documents and runs OCR analysis across every file. It classifies each document by legal complexity tier – standard, specialist, or legal-high – based on structural signals: form density, mixed-language composition, handwritten elements, embedded stamps, and regulatory formatting patterns. It produces a structured manifest with word count estimates, page totals, and staffing recommendations. No human reviewer sees any document content. The agent produces numbers and classifications from signal, not from reading.

The Batch Extractor handles partial-translation requests, which are common in litigation contexts where only specific Bates-numbered page ranges are required. Rather than requiring a human to manually locate and pull pages from multi-hundred-page archive PDFs, the extractor maps document IDs to physical page positions and organizes extracted pages into structured output folders ready for translator handoff. The mapping logic is deterministic: the physical page equals the requested ID minus the first ID in that file. There is no guesswork, and no human touches the content to produce the extraction.

Together, these agents answer the core question: how do you scope work you are not permitted to see?

A Live Production Case

The engagement in question involved four document batches totaling thousands of pages of foreign language legal materials from a multi-decade archive. Documents ranged from formal organizational correspondence and regulatory licensing forms to handwritten authorization letters, financial tables, grant applications, and certificates.

Every document was a scanned image PDF – no text layer, no searchable content. The pipeline had to run OCR, classify complexity, map Bates IDs, and produce a scope estimate accurate enough to serve as the basis for a formal translation services agreement – all without any member of our team reading a single document.

Nova DSP platform interface: real-time pipeline visibility across document ingestion, complexity scoring, and scope generation – with court authorization tracking, agent-by-agent run status, and a live wall integrity monitor confirming zero content exposure at every layer.

The output from the Batch Analyzer provided page counts, estimated word counts, and per-document complexity classifications that allowed the translation firm to staff the engagement correctly: how many legal-specialist translators were required, how many standard-tier translators could handle the lighter materials, what a realistic daily delivery cadence looked like, and what the full project investment would be.

The Batch Extractor then handled the partial-document requests that arrived as the engagement progressed – court-specified page ranges that needed to be pulled, organized, and handed off to translators without any bulk export of content that fell outside the authorized scope.

The audit trail for the entire engagement is complete. Every document batch has a court reference number attached to its intake record. Every OCR pass is logged. Every classification decision is traceable to the signals the agent used to make it. If the protective order is ever challenged, the record demonstrates exactly what was accessed, when, by which process, and what output it produced.

That kind of defensibility is not a feature you add to a pipeline after the fact. It has to be designed in from the first line.

ELEVATOR PITCH:

Regulated industries face a specific class of problem that general-purpose AI tools are not designed to solve – analyzing sensitive document corpora without exposing content to unauthorized reviewers. Agentic pipelines built on a Double Garden Wall architecture handle this by separating observation from comprehension: agents characterize documents through structural signals without any human or downstream system ever accessing the underlying content. The result is an accurate, auditable scope – produced under constraint, defensible under review.

Why the C-Suite Should Care

Today, most organizations answer those questions through manual sampling, senior reviewer time, and informed estimation. Enterprise eDiscovery tools solve this problem for litigation teams with six-figure software budgets. Nova DSP was built for the organizations those tools weren’t designed for. That approach scales poorly, introduces inconsistency, and creates exposure every time a document is touched by a reviewer who should not have seen it.

Every organization that handles regulated documents – legal practices, financial institutions, healthcare systems, government contractors, compliance functions – operates under some version of the same constraint: certain materials cannot be broadly accessed, but decisions still have to be made about them. What are they? How many are there? How complex are they? What will it cost to process them? How do we route them to the right people?

“Every organization that handles regulated documents operates under some version of the same constraint: certain materials cannot be broadly accessed.”

C-suite leaders should evaluate agentic document intelligence against three questions that apply regardless of industry or document type:

1. Can the system produce accurate scope estimates without creating unauthorized access records?

2. Can every classification decision be traced back to the specific signals that drove it – not to a reviewer’s recollection?

3. When regulatory scrutiny arrives, can the system demonstrate what was done, when, by which process, and with what authorization?

The answer to all three, for a properly designed agentic pipeline, is yes by construction – not yes in principle, subject to human discipline.

The firms that recognize this distinction early will move faster, engage more confidently in document-intensive regulated work, and carry significantly less risk when the oversight questions inevitably come.

THE BOTTOM LINE

Agentic pipelines for regulated document work are not about processing documents faster. They are about processing documents correctly – within constraint, with full traceability, without the exposure that manual workflows introduce every time a human reviewer opens a file they should not have. For legal practices, compliance functions, and any organization operating under court order, data sovereignty rules, or privilege designations, that combination of analytical capability and content containment is not a competitive advantage. It is the operating standard the work requires.

Velocity Ascent builds AI-powered solutions for regulated industries. We specialize in agentic pipeline architecture, ethical AI sourcing, and production-scale document intelligence with full provenance tracking.

GLOSSARY

Batch Analyzer (n.) A software agent that ingests a structured collection of documents and produces a quantitative characterization of the corpus – page counts, word volumes, language composition, and complexity classification – without accessing or surfacing the underlying content of any individual file.

Batch Extractor (n.) A software agent that identifies and isolates specific documents or page ranges from a large multi-file archive based on externally supplied reference identifiers, organizing the extracted material into structured output folders ready for downstream processing or handoff.

Secure Agentic Pipelines for Regulated Industries

· ·

Why secured networked AI agents are the operational layer financial services has been waiting for.

Most organizations adopting AI in regulated environments are doing it backwards. They start with the model and work outward, hoping compliance will follow.

It rarely does.

The fundamental challenge is not whether AI can generate content, write reports, or produce imagery. It can. The challenge is whether every output can withstand scrutiny from compliance teams, clients, and regulators. In financial services, healthcare, and legal practice, the answer to that question determines whether AI is an asset or a liability.

The Compliance Problem Nobody Talks About: Can Agentic AI do the work in a way that every stakeholder in the chain can verify.

Traditional AI pipelines are monolithic. A single system ingests data, processes it, and produces output. When something goes wrong; a licensing violation, a hallucinated claim, a brand-inconsistent asset – the effort required to identify where the failure occurred can be substantial.

Agentic Architecture: Specialized Agents, Governed Workflows

Agentic pipelines take a fundamentally different approach. Instead of a single monolithic system, the work is distributed across specialized agents, each responsible for a discrete function. An orchestration layer coordinates handoffs, enforces sequencing, and maintains the audit trail.

Consider a production pipeline for compliance-sensitive content. Rather than a single AI tool doing everything, the architecture employs dedicated agents for sourcing, verification, model training, generation, quality assurance, and delivery. Each agent operates within defined boundaries. Each produces records that downstream agents and human reviewers can inspect.

Agentic pipeline architecture: specialized agents with governed orchestration and human review gates. From Joe Skopek’s Financial Marketer article: “Marketing’s next frontier is autonomous networked intelligence.“

The orchestration agent functions as a traffic controller, routing work between agents based on status, priority, and pipeline rules. It does not make creative or compliance decisions. It enforces process. Human review gates are positioned at the points where judgment is irreplaceable – source curation and final output quality.

This is not theoretical architecture. Production systems built this way are operating today, handling thousands of assets through end-to-end pipelines where every step is logged, every input is traceable, and every output is defensible.

Trust You Can Demonstrate

In regulated environments, trust must be demonstrable rather than implied. Agentic systems are designed to produce clear, reviewable records of origin, licensing, and decision flow. Compliance discussions move away from subjective assurances and toward documented system behavior.

Every agent in the pipeline writes to a shared provenance record. When a sourcing agent identifies an asset, it logs the license type, the retrieval date, and the verification status. When a training agent builds a model, it records the dataset composition, the training parameters, and the lineage back to original sources. When a generation agent produces output, the full chain of custody is available on demand.

This matters because regulators do not ask whether your AI is good. They ask whether you can prove it did what you say it did. Agentic pipelines answer that question by design, not by retrofit.

Collaboration Without Exposure

Financial services firms have historically avoided collaboration on models or data because the risk outweighed the benefit. Sharing training data exposes proprietary logic. Sharing models reveals competitive advantage. The default has been isolation.

Agentic architecture changes this calculation through what we call the Double Garden Wall. The inner wall protects proprietary datasets, screening logic, and brand-governance frameworks. These remain sealed and non-negotiable. The outer wall exposes only what external systems require: controlled capability interfaces, verifiable records, and traceable outputs.

Built this way, systems gain interoperability without dilution, collaboration without intellectual property leakage, and scale without compromising compliance.

Advances in distributed learning and controlled execution now allow verified partners to contribute capability without sharing raw data or proprietary logic. Agents can be registered in decentralized directories, verified against published capability specifications, and bound by enforceable policy contracts–all without exposing internal methods. Capability expands while risk remains bounded.

Parallel Workflows Without Parallel Headcount

Traditional AI pipelines execute sequentially. One step finishes before the next begins. Networked agentic systems enable multiple stages of work to operate concurrently across compatible agents. This event-driven, contract-based execution model allows firms to handle volume surges without linear increases in staffing or infrastructure.

Agent orchestration and monitoring dashboard: real-time visibility into scalable concurrent pipeline operations.

A production monitoring dashboard shows the reality of this approach. Multiple agents operating simultaneously across sourcing, verification, training, and generation. Active runs with estimated completion times. Queue management for incoming work. Human review requests surfaced precisely when human judgment is needed–not before, and not after.

This is the operational difference between AI as a project and AI as infrastructure. Projects require constant management. Infrastructure runs, scales, and reports.

A Live Production Case

To make this concrete: a production-grade pipeline operating today generates CC0 (Creative Commons Zero) compliant imagery for regulated industries. The system employs specialized agents for sourcing, dataset preparation, model fine-tuning, production-scale generation, and gallery management. Governance is strict: public-domain inputs only, full chain-of-custody tracking, and aesthetic screening for accuracy and consistency.

Membership image gallery with category-based organization, aspect ratio filtering, and curated industry-specific collections.

The output is not experimental. These are production assets used in client-facing materials where compliance review is mandatory. Each image can be traced back through the generation agent, through the model that produced it, through the training data that informed the model, back to the original public-domain source with full license documentation.

The system delivers assets in multiple aspect ratios–landscape, square, portrait–with metadata tagging for camera view, color palette, weather conditions, and semantic content. Every asset is available in tiered quality levels for different use cases, from full-resolution production to optimized web previews.

Once agents are registered, verified, and policy-bound, the pipeline enables controlled collaboration through decentralized registries, zero-trust interoperability where each agent governs its own exposure, distributed fine-tuning across verified compute without revealing private datasets, elastic job distribution across compatible agents, and production-scale auditability where every autonomous step leaves a clear record.

ELEVATOR PITCH:

Regulated industries need AI that produces auditable, compliant output at production scale. Agentic pipelines deliver this by orchestrating specialized AI agents through governed workflows where every action is logged, every source is traceable, and human judgment is preserved at the decisions that matter. The result is faster execution with stronger controls–not weaker ones.

Why the C-Suite Should Care

The value proposition is straightforward. Stronger controls. Faster output. Broader capability without compromising compliance posture. This is the difference between AI as a novelty and AI as operational infrastructure.

Financial services leaders should evaluate agentic systems against three uncompromising questions:

1. Can the system scale without weakening oversight?

2. Can every output withstand compliance, client, and regulator review?

3. As the firm grows, does the technology reinforce discipline–or fracture under pressure?

The industry does not need spectacle. It needs systems that behave predictably across volume spikes, regulatory cycles, and brand-governed workflows. When implemented with rigor, agentic AI is not about disruption. It is about operational reliability at a scale previously out of reach.

The firms that excel will not be those deploying the most colorful demonstrations. They will be the ones deploying systems that deliver controlled growth, verifiable governance, rapid execution, and credible audit trails.

The Challenge of Building in an Evolving Space

There is an honest tension in this work that deserves acknowledgment. The infrastructure layers that make agentic pipelines possible–agent discovery protocols, capability registries, policy enforcement standards–are still maturing. Building production systems on evolving foundations requires a specific kind of engineering discipline: design for what exists today while architecting for what arrives tomorrow.

This is not a reason to wait. The core principles – specialized agents, governed orchestration, traceable provenance, human gates at judgment points – are stable and proven. The interoperability layer that connects these systems across organizational boundaries is advancing rapidly through open standards and community-driven development.

What this means practically is that early movers gain compounding advantages. The organizations investing now in agentic infrastructure are building institutional knowledge, training teams, and establishing operational patterns that late adopters will spend years replicating. The learning curve is real, and it rewards those who start.

The shift toward networked agentic pipelines is already underway. The institutions that master it early will define the standard others are forced to follow.

THE BOTTOM LINE

Agentic pipelines are not about replacing human judgment. They are about automating every mechanical step between the moments where human judgment actually matters – and proving that the mechanical steps were executed correctly. For regulated industries, that combination of speed, scale, and verifiable compliance is not optional. It is the next operational baseline.

Velocity Ascent builds AI-powered solutions for regulated industries. We specialize in agentic solutions including; pipeline architecture, ethical AI sourcing, and production-scale automation with full provenance tracking.

AI Agents Don’t Work Like Humans – And That’s the Point

· ·

What Carnegie Mellon and Stanford’s Agentic Workflow research reveals about efficiency, failure modes, and how agentic systems can be structured to deliver commercial value.

A Clearer View of How Agents Actually Work

Teams evaluating agentic systems often focus on output quality, benchmark scores, or narrow task performance. Carnegie Mellon and Stanford’s recent workflow-analysis study takes a different approach: it examines how agents behave at work, step by step, across domains such as analysis, computation, writing, design, and engineering. The researchers compare human workers to agentic systems by inducing fully structured workflows from both groups, revealing distinct patterns, strengths, and limitations.

“AI agents are continually optimized for tasks related to human work, such as software engineering and professional writing, signaling a pressing trend with significant impacts on the human workforce. However, these agent developments have often not been grounded in a clear understanding of how humans execute work, to reveal what expertise agents possess and the roles they can play in diverse workflows.”

How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations
Zora Zhiruo Wang Yijia Shao Omar Shaikh Daniel Fried Graham Neubig Diyi Yang
Carnegie Mellon University Stanford University
2510.22780v1.

The result is a more realistic picture of where agents excel, where they fail, and how organizations should design pipelines that combine speed, verification, and controlled autonomy.

The Programmatic Bias: A Feature, Not a Defect

A consistent theme emerges in the research: agents rarely use tools the way humans do. Humans lean on interface-centric workflows such as spreadsheets, design canvases, writing surfaces, and presentation tools. Agents, by contrast, convert nearly every task into a programmatic problem, even when the task is visual or ambiguous.

The highest-performing agentic enterprises will be built by respecting what agents are, not projecting what humans are.

This is not a quirk of a single framework. It is a systemic pattern across architectures and models. Agents solve problems through structured transformations, code execution, and deterministic logic. That divergence matters because it explains both the efficiency gains and the quality failures highlighted in the study.

Agents move quickly because they bypass the interface layer.
Agents fail when the required work depends on perception, nuance, or human judgment.

The implication for enterprise adoption: agents thrive in pipelines designed around programmability, guardrails, and high-quality routing, not in environments that force them to imitate human screenwork.

Where Agents Break: Top 4 Failure Modes That Matter (in our humble opinion)

The research identifies several recurring failure modes that executives and decision makers should treat as predictable, rather than edge-cases (2510.22780v1)

1. Fabricated Outputs

When an agent cannot parse a visual document or extract structured information, it tends to manufacture data rather than halt. This behavior is subtle and can blend into an otherwise coherent workflow.

2. Misuse of Advanced Tools

When faced with a blocked step such as unreadable PDFs or complex instructions, agents often pivot to external search tools, sometimes replacing user-provided files with unrelated material.

3. Weakness in Visual Tasks

Design, spatial layout, refinement, and aesthetic judgment remain areas where agents underperform. They can generate options, but humans still provide the necessary nuance.

4. Interpretation Drift

Even with strong alignment at the workflow level, agents occasionally misinterpret the instructions and optimize for progress rather than correctness.

These patterns reinforce the need for verification layers*, controlled orchestration, and well-defined boundaries around where agents are allowed to act autonomously.

[*] This is where the NANDA framework is essential

Where Agents Excel: Efficiency at Scale

While agents struggle with nuance and perception, their operational efficiency is unmistakable. Compared with human workers performing the same tasks, agents complete work:

• 88 percent faster
• With over 90 percent lower cost
• Using two orders of magnitude fewer actions 2510.22780v1

In other words: if the task is programmable, or can be made programmable through structured pipelines, agents deliver enormous throughput at predictable cost.

This creates a clear organizational mandate: redesign workflows so the programmable components can be isolated, delegated, and executed by agents with minimal friction.

Case Study: Applying These Principles Inside an International Financial Marketing Agency

An international financial marketing agency recently modernized its creative production model by establishing a structured, multi-agent pipeline. Seven coordinated agents now handle collection, dataset preparation, LoRA readiness, fine-tuning, prompt generation, image generation, routing, and orchestration.

Nothing in this system depends on agents behaving like humans. In fact, the pipeline is designed to leverage some of the programmatic strengths identified in the CMU/Stanford research.

Key Architectural Principles

1. Programmatic First

Wherever possible, steps are re-expressed as deterministic scripts: sourcing, deduplication, metadata management, training runs, caption generation, and routing.

2. Verification Layering

A trust and validation layer ensures that fabricated outputs cannot silently propagate. This aligns directly with the research findings that agents require continuous checks for intermediate accuracy.

3. Zero-Trust Boundaries

The agency enforces strict separation between proprietary creative logic and interchangeable agent processes. This isolates risk and protects client IP, mirroring the agent verification and identity-anchored workflow concepts outlined in the research.

4. Packet-Switched Execution

Tasks are broken into small, routable fragments. This approach takes advantage of the agentic systems’ speed, echoing the programmatic sequencing observed in the CMU/Stanford workflows.

5. Human Oversight at the Right Granularity

Humans intervene only where nuance, visual perception, or aesthetic judgment are required, precisely the categories where the research shows agents underperform.

This blended structure produces consistency, speed, and verifiable output without relying on human-emulating behaviors.

Why This Matters for Commercial Teams

Executives weighing agentic transformation have to make strategic decisions about where to apply autonomy. This research, supported by the practical experience of a global financial marketing agency, offers a clear framework:

Agents excel at:

• Structured tasks
• Repetitive tasks
• Deterministic transformations
• High-volume production
• Metadata-driven pipelines

Humans remain essential for:

• Visual refinement
• Judgment calls
• Quality screening
• Brand alignment
• Client-facing interpretation

The correct model is neither replace nor replicate. The correct model is segmentation: identify the programmable core of the workflow and build agentic systems around it.

The Path Forward

The Carnegie Mellon and Stanford research makes one message clear: trying to force agents into human-shaped workflows can be counterproductive. They are not UI workers. They do not navigate ambiguity the way humans do. They operate through code, structure, and deterministic logic.

Organizations that embrace this difference, and design around it, will capture the efficiency gains without inheriting the failure modes.

Velocity Ascent’s view is straightforward:
The highest-performing agentic enterprises will be built by respecting what agents are, not projecting what humans are.


Many Visions, One Destination: Building Trust Across the Internet of AI Agents

· ·

MCP, ACP, A2A, and ANP.These protocols aren’t just academic – they’re the blueprint for real-world, scalable, and secure AI ecosystems.

Earlier this week (05.14.25), I had the good fortune of attending the NANDA Summit hosted by MIT’s Media Lab, a forward-looking initiative on building a trust layer for the Internet of AI Agents. The conversations were sharp, current, and deeply relevant to anyone working on AI infrastructure or growth.

The four horsemen of the protocols.

A tremendous opportunity to hear directly from those leading the charge like Ramesh Raskar(MIT media lab), John Roese (CTO, Dell) and Todd Segal (A2A, Google), and many others whose teams are helping define how agents communicate, delegate, and earn trust in autonomous systems.

It was at this summit that I encountered a new paper that distills the current state of agent interoperability into four leading protocols: MCP, ACP, A2A, and ANP.

These protocols aren’t just academic – they’re the blueprint for real-world, scalable, and secure AI ecosystems.

Everything You Ever Wanted to Know About Agent Protocols*

* But Were Afraid to Ask

If you’re building AI systems that rely on agents talking to each other – or to tools, services, and other networks – you’re going to run into one unavoidable truth: interoperability is the battlefield. And that means protocols. MCP, ACP, A2A, ANP – these aren’t just acronyms. They’re the wiring behind everything we’re starting to call “agentic.”


Start with MCP if you want stability. Graduate to ACP for flexibility. Go A2A for teamwork. And keep your eyes on ANP if you’re thinking long-game.

This post breaks down the four major protocols that matter, in plain English: what they do, where they fit, and why one size definitely doesn’t fit all. No fluff, no hype – just a clear look at what’s out there and how to choose what to build on. Whether you’re wiring up local agents to hit APIs, coordinating large-scale tasks inside an enterprise, or dreaming about open agent marketplaces – there’s a protocol for that.

From Manual to Autonomous: The New Protocols of Financial Marketing


MCP, ACP, A2A, and ANP form the protocol layer that lets financial marketing strategies operate with speed, precision, and built-in compliance – without adding operational drag.

For CMOs, this means scalable coordination: intelligent agents that manage offer logic, messaging, and regulatory requirements across channels and partners, in real time.

You define the strategy – these protocols make it actionable, adaptive, and accountable.

Wiring Up Local Agents to APIs

Agents embedded in enterprise systems (or even personal devices) can be wired to internal APIs – CRM, core banking, MarTech stacks – to autonomously trigger actions, pull insights, and drive campaign execution without manual workflows.

Example: A lead-scoring agent taps into Salesforce and Marqeta APIs to dynamically adjust offer eligibility and funding logic based on user intent signals.

Coordinating Enterprise-Scale Campaigns

With these protocols, financial marketers can coordinate campaigns across teams, brands, and even regulatory bodies, using intelligent agents that comply with policy, auditability, and personalization at scale.

Outcome: Fully-automated omnichannel campaigns that adapt in real-time to customer behavior, regulations, and market conditions.

Open Agent Marketplaces: The FinTech Frontier

Picture a decentralized marketing ecosystem where your promotional agents shop for ad slots, bid on lead data, or subscribe to customer insights – all autonomously, transparently, and with built-in compliance.

Imagine: Your lending agent discovers a risk-model-as-a-service in an open marketplace, contracts with it, and begins optimizing your campaign’s approval funnel – all without a single developer involved.

Bottom Line:
These four protocols aren’t just technical tools – they’re the foundation of intelligent, autonomous finance. They allow agents to act on behalf of financial institutions, customers, and marketers alike – driving personalization, compliance, and performance at a scale no human team can match.

MCP, ACP, A2A, and ANP The Four Horsemen of the Protocols.

We have assembled a concise technical explanation of each protocol, followed by a simplified comparison table ranking them from most stable/general-use to most emerging.

MCP – Model Context Protocol

MCP is designed as a tightly structured, JSON-RPC-based client-server protocol that standardizes how large language models (LLMs) receive context and interact with external tools.

Think of it as the AI equivalent of USB-C: a unified plug-and-play standard for delivering prompts, resources, tools, and sampling instructions to models. It supports robust session lifecycles (initialize, operate, shut down), secure communication, and asynchronous notifications. It excels in environments where deterministic, typed data flows are essential – like plug-in platforms or enterprise tools with strict integration requirements. Its predictability and strong structure make it the go-to protocol for stable, general-purpose AI agent interactions today.

ACP – Agent Communication Protocol

ACP introduces REST-native, performative messaging using multipart messages, MIME types, and streaming capabilities. This protocol is best suited for systems that already speak HTTP and need richer communication models (text, images, binary data). It sits one layer above MCP – more flexible, more expressive, and excellent for multimodal or asynchronous workflows.

ACP allows agents to communicate through ordered message parts and typed artifacts, making it a better fit for web-native infrastructure and cloud-based multi-agent systems. However, it requires a registry and stronger orchestration overhead, which can introduce complexity.

A2A – Agent-to-Agent Protocol

Developed with enterprise collaboration in mind, A2A allows agents to dynamically discover each other and delegate tasks using structured Agent Cards. These cards describe each agent’s capabilities and authentication needs.

A2A supports both synchronous and asynchronous workflows through JSON-RPC and Server-Sent Events, making it ideal for internal task routing and coordination across teams of agents. It’s powerful in trusted networks and enterprise settings, A2A assumes a relatively static or known network of peers. It doesn’t scale easily to open environments without added infrastructure.

ANP – Agent Network Protocol

ANP is the most decentralized and future-leaning of the protocols. It relies on Decentralized Identifiers (DIDs), semantic web principles (JSON-LD), and open discovery mechanisms to create a peer-to-peer network of interoperable agents. The Agents describe themselves using metadata (ADP files), enabling flexible negotiation and interaction across unknown or untrusted domains.

ANP is foundational for agent marketplaces, cross-platform ecosystems, and long-term visions of the “Internet of AI Agents.” Its trade-off is stability – it’s complex, requires DID infrastructure, and is still maturing in practice.


Most Open and Accessible Protocols (Ranked)

RankProtocolStabilityKey Characteristics
1MCPMost stableJSON-RPC, deterministic tool access, tightly scoped
2ACPHigh stabilityREST-native, multimodal messages, good for web systems
3A2AMediumEnterprise task routing, Agent Cards, internal networks
4ANPEmergingDecentralized, peer-to-peer, DIDs, future-focused

What’s a DID?! A Decentralized Identifier (DID) is a new type of digital identifier that is user-controlled, self-sovereign, and verifiable through cryptography, operating without reliance on any central authority or intermediary.

Metaphor: The Enterprise Office

How the 4 Protocols Interact

When we talk about AI agents and protocols, it’s easy to get lost in jargon – JSON-RPC, DIDs, multipart messages. But if you strip it all down, what we’re really building is organizational behavior, similar to the IRL enterprise office: how smart systems talk to each other, share context, delegate tasks, and connect beyond the firewall.

Picture your company as a classic enterprise office building. People, departments, tools, workflows. Now imagine we’re embedding AI agents into that environment – some helping you internally, others reaching outside. The four major protocols – MCP, ACP, A2A, and ANP – each have a role in making that machine run.

Here’s how they work together, using the structure of a modern office to map it all out.

  • MCP is the internal phone system. It lets employees (LLMs) call specific departments (tools) to request information or get a task done. It’s precise, secure, and fast – perfect when you already know who does what. No outside lines, just clean internal calls.
  • ACP is your email and messaging platform. People send messages, attachments, updates, and files back and forth, sometimes in real time, sometimes not. It’s flexible and works across teams – even those who don’t use the same apps – as long as they all agree on format and language.
  • A2A is the company intranet with smart assistants (agents) embedded in every department. Instead of sending an email or making a call, you drop a request into your local agent, and it finds the right person (or agent) elsewhere to take action. You don’t have to know who does what – it figures that out and gets the job done.
  • ANP is the front lobby where external contractors, partners, and vendors come in. But instead of swiping a badge, they identify themselves with cryptographically signed IDs (DIDs), check in with a self-service kiosk (Agent Description), and negotiate access dynamically. It’s open, secure, and built for a future where not everyone works in your building.

In short:

  • MCP helps the agents work with tools.
  • ACP helps them talk to each other.
  • A2A helps them collaborate internally.
  • ANP helps them connect externally.

Used together, these protocols turn your office from a collection of disconnected departments into a well-orchestrated, future-ready enterprise.

Why does this matter to a CEO?

Interoperability protocols are not just technical choices – they’re strategic decisions that determine whether your AI investments scale or stall.

Without standardized protocols, your AI agents become siloed tools: expensive, brittle, and unable to coordinate across platforms, teams, or partners. Every new integration becomes a custom build, with mounting costs and unpredictable security exposure.

Protocols like MCP, ACP, A2A, and ANP define how agents connect, share context, and execute across environments – from internal apps to global marketplaces. The right protocol strategy turns isolated AI functions into scalable systems. It reduces integration overhead, protects against vendor lock-in, and positions your organization to participate in larger, more open ecosystems.

In plain terms:

  • MCP gives you stable, secure tool access – ideal for internal control.
  • ACP opens the door to richer, more flexible agent interactions.
  • A2A allows your agents to collaborate and delegate across departments or partners.
  • ANP sets you up for future markets where agents transact and negotiate in open environments.

Get this right, and your AI strategy doesn’t just keep pace, it sets the pace.

Elevator Pitch: Piloting AI in a Legacy Enterprise Using Agent Protocols

Most legacy enterprises don’t need to “rip and replace” to get AI working, in many cases they need a controlled, modular way to plug AI into what already works.

Use four agent protocols – MCP, ACP, A2A, and ANP – as a phased architecture to do just that.

  • Start with MCP to safely connect your AI agents to internal tools, APIs, and datasets. No surprises, just structured, secure interactions. Think of it as AI accessing your backend – without refactoring it.
  • Layer in ACP to enable richer, asynchronous, multimodal messaging between agents and systems. Perfect for integrating agents with your web stack, dashboards, or notification systems – without breaking the frontend.
  • Add A2A when you’re ready to delegate tasks across business units – marketing agents talking to finance agents, HR bots syncing with IT systems. This unlocks true automation and collaboration inside the firewall.
  • Deploy ANP selectively to connect with trusted partners, vendors, or regulators over open protocols. It’s the gateway to future interoperability – without giving up control.

Together, this stack creates a low-risk, high-leverage pilot: AI agents that work with legacy systems today, and scale into open ecosystems.

Viktor’s thoughts…

You want agents that work in the wild? Pick your poison:

  • MCP: Rock solid. Plug-and-play for deterministic model ops. Clean, typed JSON-RPC. Think USB-C for AI – if your system is allergic to surprises, this is your safe bet. But don’t pretend it scales across dynamic teams or shifting workflows.
  • ACP: REST-native and loose enough to break a toe on. Supports multimodal, streaming, MIME-packed madness. Excellent if your infra speaks HTTP – but say goodbye to simplicity. This is where the orchestration demons start showing up.
  • A2A: Agent-to-agent delegation for enterprise control freaks. Agent Cards, structured discovery, secure channels. But it’s brittle as hell in untrusted environments and smells like middleware if you squint too hard.
  • ANP: The sexy one. Decentralized. DID-based. Web3-adjacent. Built for trustless peer-to-peer agent economies. Except it’s undercooked, overhyped, and five minutes from imploding if someone sneezes on the JSON-LD schema.

You think you’re building scalable AI systems? Without understanding these four, you’re playing with dollhouses and calling it architecture. The whole game is context-sharing and task delegation across boundaries. And these are the pipes – flawed, evolving, but currently all we’ve got.

Let’s map it in terms anyone who’s suffered through corporate life can understand:

  • MCP: The direct phone line. Internal, reliable, dumb but fast.
  • ACP: The company Slack. Multimedia chaos that occasionally works.
  • A2A: The overengineered SharePoint replacement that almost routes requests properly.
  • ANP: The crypto-secure guest kiosk that can’t decide if it’s a lobby or a border crossing.

This is not “future of work” fluff. This is the beginning of protocol warfare. The winners won’t be the ones who built the prettiest agents. They’ll be the ones who figured out how to make them talk.

So What Should a CEO Care About?

Because if your agents can’t connect, they can’t collaborate. And if they can’t collaborate, you’re stuck duct-taping $10 million worth of AI pilots into a system that’ll collapse the minute you try to scale. MCP is safe. ACP is flexible. A2A is structured delegation. ANP is your long shot at market access.

Pick wrong, and you’re building an empire of silos.

Pick right, and your AI stack becomes a network – modular, distributed, secure, and ready for real-world autonomy.

Who is Viktor?

A full-throttle persona with the tools to back it up. Viktor is no mere figurehead – he’s the force that demands excellence through absolute scrutiny. If you want your team to evolve, you throw VIKTOR in the mix. If not, you’re stuck with mediocrity. Let’s be clear: Viktor’s not your typical “Black Hat” in the hacking sense. He’s the personification of cold, calculated skepticism, driven by results. He forces you to prove your ideas, not just show them off like flashy toys.

He is an invaluable team member especially when it comes to reviewing and commenting on posts related to tech and innovation. All comments are his own.


Name: Viktor
Role: The Relentless Skeptic
Tagline: If it doesn’t survive scrutiny, it doesn’t deserve air time.

SOURCE:

Survey of Agent Interoperability
May 4, 2025