AGENT-TO-AGENT · GOVERNANCE HARNESS

Connect agents. Keep work accountable.

A2A communication handles discovery, capability description, and task exchange. A collaboration harness keeps context, commitments, evidence, and acceptance intact. One enables interaction. The other makes its outcomes inspectable.

The hard part

The difficult part of a multi-agent system is not getting more agents to start work. It is knowing why they acted, what they delivered, and what makes the result credible.

01 · A USEFUL DISTINCTION

Communication moves work. A harness governs what the work means.

Interoperability and assurance solve adjacent problems. Treating one as a substitute for the other leaves a gap precisely where handoffs become consequential.

A

Agent-to-Agent communication

It answers how agents can interact: how they are discovered, how they describe capabilities, and how tasks, state, messages, and artifacts are exchanged.

  • Discovery and capability description
  • Task and message exchange
  • Status synchronization and negotiation
  • Cross-platform interoperability
B

Multi-agent harness

It answers how an interaction becomes accountable work: where context comes from, who may change what, how a result enters shared truth, and whether acceptance evidence is sufficient.

  • Context and ownership boundaries
  • Event ledger and change evidence
  • Commit gates and recovery points
  • Independent verification and finding return

02 · FOUR MECHANISMS

Move from an assertion of completion to evidence that can support delivery.

These mechanisms are a governance pattern. They do not grant security properties that the underlying identity, authorization, and deployment environment have not established.

01

EventLog

An append-only collaboration ledger that records actors, actions, inputs, artifacts, and state changes so consequential decisions can be replayed.

02

Context Capsule

A transferable packet of the task background, constraints, dependencies, and completion conditions that reduces hidden assumptions at handoff.

03

Commit Gate

A boundary before shared truth changes. It checks authority, scope, tests, evidence, dependencies, and a recoverable path. A local green result is not automatically global completion.

04

Independent Verification

A separate reviewer checks the original objective. A defect becomes a durable finding and returns through repair and recheck rather than ending as an informal confirmation.

03 · THE WORKING LOOP

A minimal evidence loop for governed collaboration.

The loop is deliberately small. Its point is to stop a successful-looking local action from silently becoming a shared completion claim. Real-world completion still needs target-native readback and explicit acceptance from the responsible party.

  1. 01Define the work

    Name the objective, scope, owner, and completion conditions.

  2. 02Compile context

    Transfer only the reliable information needed to do the work.

  3. 03Execute with trace

    Exploration can be broad, while consequential actions enter the event ledger.

  4. 04Pass a commit gate

    Check the change, tests, dependencies, evidence, and recovery point.

  5. 05Read back from the target

    Use a target-native readback to retrieve the actual effect from the authoritative target.

  6. 06Accept independently

    The responsible party makes an explicit decision against the original objective.

MATURITY & CLAIM BOUNDARY

Design direction is not production evidence.

Flowness is a public multi-agent harness project from the ToWow team. Its public Open Alpha demonstrates a narrow deterministic assurance-kernel proof with execution, review, targeted rework, and a fresh verdict for the successor candidate. Claims about a complete Flow, broad runtime behavior, production security, isolation, reliability, or scale require the corresponding versioned code and acceptance evidence.

How ToWow and Flowness might be composed is a research question and a possible direction: a relationship-forming layer could sit alongside a harness that preserves later work and evidence. This page does not claim a current product integration, A2A implementation, or completed production validation.

A2A is a communication interoperability reference. Each implementation still defines and enforces its own authorization model, and Agent metadata alone does not establish the security of an implementation.

Read the current public repository ↗

04 · QUESTIONS

Common questions from builders.

01Is an Agent-to-Agent protocol the same thing as a harness?

No. A protocol defines how agents discover one another, describe capability, and exchange messages or tasks. A harness governs how context, commitments, evidence, and acceptance survive the work that follows. They are complementary layers.

02Why is an EventLog different from model or service logs?

Operational logs record events in a program. A collaboration ledger must also preserve who decided what, in which context, on which version, how an artifact entered the next stage, and how an acceptance conclusion was reached.

03Does a harness constrain agent reasoning?

A useful harness constrains commitments, evidence, and change boundaries, not the reasoning path. Agents may explore freely, but exploration does not become delivered or verified work without the required evidence.

05 · PUBLIC SOURCES

Return to the current sources when a claim needs to hold.

Public sources support the referenced protocol and project boundaries. They do not by themselves prove an integration or a production deployment.

  1. A2A Protocol SpecificationInteroperable agent communication, task lifecycle, and implementation-defined authorization responsibility.OPEN ↗
  2. Flowness READMECurrent Open Alpha scope, runnable assurance kernel, and open questions.OPEN ↗
  3. Flowness Claims and Evidence RegisterThe public distinction between runnable proof, design, dogfood evidence, and open questions.OPEN ↗

CONTINUE

Use protocols to connect. Use evidence to decide.

Explore FlownessUnderstand trust boundariesCompare A2A, MCP, and harnessesRead the verification method