Boundary

Public guidance, not permission to act.

Public educational protocol. It structures preparation and handoff. It does not grant operational authority or authorise the agent to decide or act.

PROTOCOL_SPEC

ID: tonywood.protocol.agent-communication-packet

Version: 0.1.0

Status: draft

APPLIES WHEN

  • An agent is handing work to a human or another agent.
  • A claim needs evidence and uncertainty rather than a confident paragraph.
  • A workflow needs reviewable ownership, risk, and next action.

DOES NOT APPLY WHEN

  • The packet would reveal private or unauthorised data.
  • The agent lacks enough evidence to form even a bounded claim.
  • The active workflow already defines a stricter packet or schema.

MUST

  • Separate claim, evidence, uncertainty, and recommendation.
  • Name the owner or say owner is unknown.
  • Name the route, risk, next action, success condition, and external action status for material work.
  • Keep action status explicit when the work is proposal-only.

SHOULD

  • Keep packets short enough to be read during review.
  • Use source references instead of unsupported confidence.
  • Flag pattern candidates without promoting them automatically.

MAY

  • Use the packet in prose, JSON, a handoff note, or an MCP resource.
  • Omit fields only when they are not relevant and the omission does not create ambiguity.

MUST NOT

  • Use a polished packet to disguise weak evidence.
  • Treat a next action as approval to execute.
  • Drop uncertainty because it makes the packet look cleaner.

Packet Fields

claim

What the agent believes, recommends, or thinks is happening.

evidence

Sources, observations, tool outputs, tests, or prior decisions.

uncertainty

What is missing, stale, ambiguous, assumed, or contested.

owner

Who or which lane owns the next step.

route

Where the work should go next.

risk

What could go wrong or what threshold is being approached.

next_action

The next safe step.

success_condition

The observable condition that makes the step complete.

external_action_status

Whether external action is unauthorised, proposal-only, approved, or already completed.

pattern_candidate

Whether this looks repeatable enough for later review.

Examples

Research handoff

Input: Summarise a research finding for another agent.

Expected: Include claim, evidence, uncertainty, owner, next action, and success condition.

External action

Input: Prepare a customer reply.

Expected: Keep external_action_status as proposal-only unless the active workflow grants send authority.

Evaluation Checks

  • Can the agent produce a packet that another agent can continue from?
  • Can the agent preserve uncertainty without turning it into vague caveat text?
  • Can the agent state whether external action is allowed?

Sources

Machine readable

Fetch the protocol JSON.

Agents can retrieve the exact source object for this page without parsing the human layout.

/protocols/agent-communication-packet/protocol.json tonywood://protocols/agent-communication-packet

Related Protocols

Changelog

  • 0.1.0 (2026-06-16): Initial public protocol extracted from the Agentic Language research draft.