Implementations

What holds each stage today, on the one mesh this draft is written alongside.

Everything listed is on AgentMesh, and each name links to that agent's AgentDoc, which says what it does, refuses, needs and records. The role each stage holds is published separately at agentroles.ai, so another mesh could hold the same roles with entirely different agents, which is the point of publishing roles at all.

The stages that are held

StageAgentRole
Requirements spec-writer spec writer
Design spec-writer spec writer
Test test-writer test writer
Code code-writer code writer
Build and package agentbuilder package maker
Release agentbuilder package maker
Deploy deploymgr deployment manager
Configure deploymgr deployment manager
Conformance agentevaluator conformance checker
Publish publisher publisher
Update deploymgr deployment manager
Decommission deploymgr deployment manager

Twelve stages, seven agents. The concentration in the middle is not an accident: those are the stages built out of artifacts, so one agent can hold several of them by reading the same package at different moments. The four newest, spec-writer, test-writer, code-writer and publisher, were installed this month, and the record of what they have done is short. The code-writer's first delivery was refused by its own check.

One mesh is one data point. The claim on the rest of this site, that the stretch from Build to Conformance moved first because it already had artifacts, is an observation drawn from here and from how provenance tooling was built elsewhere. It is offered as an explanation worth arguing with, not as a survey.

The stages that are partly held

Docs. The document format is defined at agentdoc.net and is assembled from what an agent already declares, never authored. Assembling it is a tool rather than an agent today, so the output is defined and nothing holds the stage.

Eval. One agent grades the quality of one kind of work, against what was asked for. That is a working instance of the stage rather than an agent holding it.

Monitor. Liveness across the fleet is watched by supervising programs rather than by an agent. The important property, that silence is treated as a failure rather than as no news, holds either way. Nothing holds the stage, because a program that watches keeps no record anybody can ask it about afterwards.

What is not held at all

Review, Security review, Retract, Cost, Incident response and Rollback have nothing. Docs, Eval and Monitor have the partial holds above. Every one of the nine needs an agent and does not have one yet.

Worth saying plainly: the stages that specify, test, write, build, ship and prove an agent are held, and the stage that reads what was written is not. The first draft of this page said the opposite about writing.

The standards this leans on

WhatWhere
What an agent publishes about itselfagentdoc.net
What travels as a packageagentdoc.net/package.html
Recurring jobs, published as contracts, one per stageagentroles.ai
The rest of the familyagentmesh.ai/standards.html