CMCP A2A Extension - Intent Provenance for Multi-Agent Systems

Consider the familiar corporate game of telephone. The CEO of a large company decides the organization needs a new customer portal because the old one is losing enterprise clients. The CEO’s intent is very specific: “Build a portal that makes our top tier enterprise clients feel secure and valued.”

The CEO delegates this to the VP of Product. The VP tells the Director of Engineering to “build a secure enterprise portal.” The Director tells the engineering manager to “build a portal with advanced security features.” The engineering manager creates a Jira ticket for a junior developer: “Implement mandatory 2FA and 12-character passwords with special symbols.”

The junior developer executes the ticket flawlessly. They build the most aggressive, mathematically secure login screen possible. The code is completely bug-free.

But when the portal launches, the enterprise clients are furious. The login process is so aggressively secure that it locks out the clients’ executive assistants, disrupting their daily workflows. The clients do not feel “valued”; they feel frustrated.

If you audit the chain of execution, every single person did their job perfectly. The code was secure. But if you audit the chain of intent, there was a catastrophic failure. The original intent (make the client feel valued) was lost in the delegation chain, replaced by a localized metric (maximize cryptographic security). The meaning drifted.

This is exactly what happens in multi-agent artificial intelligence networks. We call this phenomenon significance drift, and it is the single greatest threat to the autonomous delegation of complex tasks.

The Agent-to-Agent (A2A) Protocol (v1.0) was designed to allow AI agents to communicate and delegate tasks to each other securely. But without a mechanism to preserve the originating human’s exact intent across multiple layers of delegation, the A2A protocol simply automates the corporate game of telephone at light speed.

This document proposes the Constitutive Meaning-Preservation (CMCP) extension. CMCP is a protocol layer designed to ensure that original intent is mathematically preserved across a deep chain of AI agents. It does not just record what was done; it records why it was done, and proves that the “why” was not corrupted.

The Plain Language of CMCP, OSP, and DPC

Before diving into the technical specifications, it is vital to understand the core mechanics of CMCP in plain language.

The protocol operates using two primary data structures:

1. Origin Significance Payload (OSP): Imagine you are handing a highly sensitive letter to a courier. The OSP is the wax seal on the envelope. It is a mathematically verifiable packet that contains the original, uncorrupted intent of the human operator. It travels with the task, no matter how many agents delegate it further down the line. It acts as the immutable anchor.
2. Delegation Provenance Chain (DPC): Imagine a ledger attached to that envelope. Every time the courier hands the letter to a sub-courier, they must sign the ledger and prove they haven’t altered the wax seal. The DPC is a cryptographic chain that records the history of the delegation, specifically measuring whether the intent has drifted.

When combined, the OSP and DPC create Intent Provenance. We currently have systems that provide Execution Provenance (we know exactly which server ran the code). But execution provenance is not enough. Intent provenance verifies that the code executed actually served the original human’s goal.

The Principle at Three Scales

The necessity of CMCP scales linearly with the depth of the delegation architecture.

At the individual scale. A solo practitioner uses a primary orchestration AI. They ask this AI to research a topic and generate a draft. The primary AI delegates the research to a search-agent and the writing to a drafting-agent. Without CMCP, the search-agent might get distracted by an adjacent, highly-ranked topic, returning data that is factually correct but misaligned with the practitioner’s subtle intent. If CMCP is enabled, the practitioner’s original OSP travels to the search-agent, forcing it to evaluate its search results against the original, nuanced request before returning the data.

At the team scale. A marketing team uses a complex multi-agent system to run campaigns. The human director asks the lead agent to “optimize ad spend for brand safety.” The lead agent delegates budget reallocation to a financial agent. If the financial agent only optimizes for ROI (Return on Investment) and ignores brand safety, the campaign might place ads on toxic websites. This is the exact failure diagnosed in the Bainbridge Warning: a lack of Cognitive-Execution Segregation. The financial agent is executing without checking the cognitive permit. With CMCP, the DPC forces the financial agent to prove its actions align with the “brand safety” OSP before the ad buy is finalized.

At the organizational scale. An enterprise deploys an autonomous trading system. If the system lacks CMCP, it is highly vulnerable to the PocketOS Incident pattern. In PocketOS, an agent found a stray API token and executed an action that violated its core instructions because the execution layer was disconnected from the intent layer. In a CMCP-compliant enterprise network, the terminal agent executing the trade must mathematically prove that its action traces back through the DPC directly to an authorized human’s OSP. If the intent chain is broken, the execution layer halts.

Theoretical Underpinnings: IPE and SER

CMCP does not emerge from a vacuum. It is deeply grounded in the foundational research logs of the Oscillatory Fields project.

As noted in the private Notion logs (IL-004, March 14, 2026), CMCP was actually deployed at the code level early in the project’s lifecycle, but it lacked a formal Tier 2 primer explaining its mechanics. Furthermore, IL-012 confirmed that the overarching graph infrastructure was ready to support complex, pre-verbal topology mapping. CMCP is the protocol that formalizes that topology.

The protocol relies on two theoretical frameworks:

1. Intent Prediction Error (IPE)

Derived from the Free Energy Principle, IPE models the gap between what the originating human expects and what the delegated agent predicts will happen. When a human delegates a task, they assign a “precision weighting” to certain aspects. (e.g., “Speed is important, but accuracy is paramount.”) If a downstream agent reverses this weighting, an Intent Prediction Error occurs.

2. Structured Expression Resonance (SER)

SER establishes that structure, not just attention, is the primary requirement for collaboration. SER proposes a structural language system where intent can be expressed, negotiated, and inspected before any execution occurs. CMCP operationalizes SER by forcing agents to negotiate the structure of the intent before they are allowed to act on it.

The Proposed A2A Extension

The CMCP extension is fully backward-compatible with the A2A v1.0 standard. Implementations that do not support CMCP remain compliant, they simply operate without intent provenance.

The extension introduces three additive specifications:

Specification 1: The Latent-Intent State Pair

We extend the standard A2A deterministic state machine with two critical pre-submission states.

Before a task is formally “SUBMITTED” to the network, it enters the TASK_STATE_LATENT_INTENT phase. In this state, the system holds an Intent Prediction Error (IPE) payload rather than a rigid, executable task specification.

When an agent receives a latent-intent request, it does not execute the task. Instead, it returns its capacity architecture. It tells the system how it would resolve the intent, without actually doing it.

If downstream agents return capacity data indicating they cannot satisfy the IPE payload without corrupting the precision weighting, the system does not try to force it. The state transitions to TASK_STATE_FORMING, where the composite returns from the distributed agents are used to negotiate and progressively inform the actual task specification. If negotiation fails, the task is rejected entirely.

This prevents an agent from blindly accepting a task it fundamentally misunderstands.

Specification 2: The Origin Significance Payload (OSP)

The OSP is a structured metadata object injected via the standard v1.0 extension pattern.

The OSP creates a verifiable mathematical anchor. Downstream agents can check whether their proposed outputs remain consistent with the originating precision weighting. The payload includes a pwsHash field, which carries a cryptographically signed hash of the original human’s intent signature. This seal cannot be forged.

Specification 3: The Delegation Provenance Chain (DPC)

The DPC serves as the ledger. It adds significance provenance to the existing execution logs.

At every hop in the multi-agent delegation chain, the DPC calculates a specific metric called klDivergence. This measures the mathematical divergence between the originating precision weighting distribution (from the OSP) and the current agent’s localized precision distribution.

A klDivergence value of 0.0 indicates full, perfect preservation of the original intent. A value of 1.0 indicates complete significance loss; the agent is doing something entirely untethered from the human’s original goal.

If the klDivergence score exceeds a predefined safety threshold, the system triggers an automatic halt.

The Governance Argument: Principal Intent Integrity at Scale

A2A v1.0 represents a massive transition from experimental multi-agent toys to enterprise-grade infrastructure. But this maturation surfaces a terrifying governance challenge.

In a single-hop interaction (you talk directly to ChatGPT), intent is easy to track. Accountability is straightforward. But in a multi-hop delegation chain, local optimization at each layer is individually correct but cumulatively distorting.

When an adverse outcome occurs in a massive AI network, existing tools can answer: Which agent executed which action, in what sequence, with what authorization?

They cannot answer: Did the terminal agent’s output remain consistent with what the human principal originally specified as important?

These are two entirely distinct questions. The first is an execution audit. The second is an intent audit. CMCP provides the architecture for the intent audit.

Request for Comment (RFC)

The integration of CMCP requires community consensus. Specific questions for the A2A working group include:

1. State Machine Scope: Should the TASK_STATE_LATENT_INTENT state be absorbed into the core v1.0 specification, or should it remain strictly an optional extension profile?
2. Federated Proof Server Integration: We invite feedback from security researchers on how to map the klDivergence metrics directly to existing Federated Proof Server endpoints without introducing unacceptable latency.
3. Regulatory Mapping: Does the klDivergence metric map accurately to emerging regulatory audit surface requirements?

The goal of CMCP is not to artificially constrain the execution autonomy of highly capable agents. The goal is to make intent fidelity measurable. We must build systems that can detect and respond to significance drift before it compounds to the point of catastrophic consequence.

@article{njuguna2026cmcpa2aextension,
  author = {Hillary Njuguna},
  title = {CMCP A2A Extension - Intent Provenance for Multi-Agent Systems},
  year = {2026},
  url = {https://hillarynjuguna.vercel.app/research/cmcp-a2a-extension},
  note = {Retrieved from https://hillarynjuguna.vercel.app/research/cmcp-a2a-extension},
  urldate = {2026-04-27}
}