Join the lifecycle where you are.

Every lifecycle runs on its assets: the statement, the tests, the code, the package, the records. Everybody starts with some of them already.

Start from the furthest thing you already have.

  1. If you already have an idea,start at RequirementsWrite the statement of the job. Everything after is checked against it.
  2. If you already have a prototype,start at Test, then CodeWrite the statement of the job from your prototype, then Test.
  3. If you already have working code,start at Package on-rampBring the statement and your programs; Package lists every program.
  4. If you already have a package or an image,start at Release on-ramp, then Deploy goes liveGive it an address and a signed build record, then deploy it.
  5. If you already have a remote agent with an A2A endpoint,register it with a platformIts Agent Card gives most of its Descriptor. Add what it refuses and what records it keeps. No package needed.
  6. If you already have an agent connected to a network through a node,it's already on; start at DocsFinish its Descriptor so people can find it.
  7. If you already have agents running on a platform,start at Run and ChangeWatch them, judge their work, and change them with a record.

Every asset fills in a little more of the picture.

Here they are in stage order. Dashed ones are optional. Solid ones are required. Every agent has a signed Descriptor and its registration, and every agent a platform runs also has the package and the signed records from Release on, each as it happens.

  1. Statement of the jobRequirementsoptional
  2. Architecture recordArchitecture and Designoptional
  3. DesignArchitecture and Designoptional
  4. Impact reportImpactoptional
  5. Catalog answersSourceoptional
  6. Test suiteTestoptional
  7. CodeCodeoptional
  8. Review findingsReviewoptional
  9. Security statementSecurity reviewoptional
  10. The packagePackage on-rampplatform-run
  11. Signed build recordRelease on-rampplatform-run
  12. Signed Descriptor and registrationDeploy goes liveevery agent
  13. Signed deployment recordDeploy goes liveplatform-run
  14. Conformance verdictConformanceplatform-run
  15. AttestationsConformanceplatform-run
  16. AcceptanceAcceptanceplatform-run
  17. AgentDocDocsplatform-run
  18. Published addressPublishplatform-run
  19. Retraction recordRetractplatform-run
  20. Receipts, usage and costCostplatform-run
  21. Graded judgmentsEvalplatform-run
  22. Incident recordsIncident responseplatform-run
  23. Change recordsUpdate, Reconfigure, Rollbackplatform-run
  24. Decommission recordDecommissionplatform-run
  • optional: whatever your lifecycle makes before the package
  • every agent: the signed Descriptor and its registration
  • platform-run: every agent a platform runs has the package, and the records from Release on, as each happens

Here is every asset, its parts, and the stage that makes it.

AssetIts partsMade atWho has it
Statement of the job
  • what it does
  • what it refuses
  • what it needs
  • what it records
  • what it costs
  • the questions only the person asking can answer
Requirementsoptional
Architecture record
  • the systems it reaches, in plain words
  • the product bound to each, in dev, test and production
  • dependencies on other agents and services
  • the agent's kind
  • limits on cost, speed, data and trust
  • each decision with its reason
Architecture and Designoptional
Design
  • the steps, in order
  • what each step is given and hands back
  • which steps are program and which are judgment
Architecture and Designoptional
Impact report
  • each dependent
  • each marked compatible, breaking or a change of behavior
Impactoptional
Catalog answers
  • for each step, an agent or solution that fits, or "build one", with the reason
Sourceoptional
Test suite
  • acceptance tests
  • unit tests
  • refusal tests that check it says no when it should
  • what no test could be written for
Testoptional
Code
  • the programs
  • an account of what was written
  • the record of every round
Codeoptional
Review findings
  • each naming a place, a claim and a consequence
  • the answer to each
Reviewoptional
Security statement
  • what it may reach
  • what it may spend
  • what it must never do
  • the reason for each permission it asks for
Security reviewoptional
The package
  • the draft Descriptor (agent-descriptor.json)
  • the runtime requirements (agent-runtime-requirements.json)
  • skills and scripts
  • a manifest
  • a digest
Package on-rampplatform-run
Signed build record
  • who built it, from what
  • the package at an address, under its digest
Release on-rampplatform-run
Signed Descriptor and registration
  • the Descriptor, bound to the agent's identity and signed
  • its registration, pointing at the current Descriptor
  • for an agent running elsewhere, made by its operator
drafted at Package on-ramp, bound and signed at Deploy goes liveevery agent
Signed deployment record
  • the package digest
  • credentials placed by name
  • ceilings set
  • the checks run
Deploy goes liveplatform-run
Conformance verdict
  • signed, case by case, against the Descriptor
Conformanceplatform-run
Attestations
  • one signed statement per version, per attester
  • the check list it used, by address
  • each check passed, failed, or not attested, which is not the same as failed
  • who attested, and whether that was the agent's own owner
  • stale when a new version ships, or after a set time
Conformance, and by any attester after: see your checking toolplatform-run
Acceptance
  • the requester's recorded yes against "done", or a no that says what is missing
Acceptanceplatform-run
AgentDoc
  • what the agent does, refuses, needs and records, each claim traced to its source
Docsplatform-run
Published address
  • an address a person can open
  • the digest of what is there
Publishplatform-run
Retraction record
  • what was taken down
  • by whom, and when
  • why
Retractplatform-run
Receipts, usage and cost
  • what was spent, by whom
  • against what was allowed
Costplatform-run
Graded judgments
  • a grade on each piece of work, with its reason
  • against a standard set beforehand
Evalplatform-run
Incident records
  • a timeline of what is known
  • every fact with its source
  • every gap, marked
Incident responseplatform-run
Change records
  • what changed, signed
  • for an update, the candidate run before the promote
  • for a rollback, the reason and what can't be undone
Update
Reconfigure
Rollback
platform-run
Decommission record
  • nothing running
  • the records kept
  • the address retired
Decommissionplatform-run

Only the ending is fixed.

Everything before the package is yours to shape. Every agent, wherever it runs, ends with a signed Descriptor and a registration that points at it. An agent a platform runs also has its Agent Package and the records each later stage leaves.