Skip to main content

The Life of an Agent

An agent is not conjured from a prompt. It is built, admitted, given an identity, equipped, briefed, and put to work, then coordinated, governed, watched, paid for, and eventually retired. Twelve steps, and the point of this page is that every one of them is a real surface, not a stage in a pitch.

The arc is the same one a member of staff moves through, and it is deliberate: an agent is an entity, not a key, so it earns an identity, carries a record, and costs money like any other actor in the estate.

The arc​

Figure D1: the life of an agent. Mint issues a signed identity (green); admit and brief are in-flight gated steps (cyan); govern is the Net Potential Gain (NPG) gate that can refuse an action (amber). Each box names the surface that performs the step. The behavioural scoring behind the govern step is on the roadmap, the refusal gate itself is what is live (see the ceiling below).

Table equivalent

#StepThe real surfaceWhat happens
1AuthorThe Studio: Mint Agent wizard, blueprint/skill/function editors, formation composerAn agent is built: a blueprint, its skills, its functions, the shape of its team
2AdmitAdmission (policy gate, with a preview of the verdict)The entity is admitted against policy before it exists as an actor; an operator sees the verdict first
3MintMint Agent wizard → cell certificate authorityEd25519 principal material plus an X.509 certificate carrying its ID and clearance; identity ≠ authority (envelope starts empty)
4EquipSkill catalog, function endpoints, tool/MCP catalogTools, skills, and functions are attached: each permission-gated and journaled
5BriefField Manuals + briefing reader + onboarding checklistA field manual is resolved for the entity and an acknowledgement is recorded before work is assigned
6TeamFormation composer, billet editor, Instantiate Crew wizardThe agent is placed in a formation; a billet is a post, a crew is the staffed formation
7RunThe Agent RuntimeDurable, signed, externally replayable execution: the plan is persisted, not held in the chat
8CoordinateTiered blackboard, consensus engine, messagingPeer agents claim and complete tasks with no manager relay; conversations ride the governed messaging fabric
9GovernThe Net Potential Gain (NPG) gateEvery consequential action is checked at the boundary and can be refused: on depth, fan-out, or net effect
10ObserveCrew monitor, run watch, management treeThe work zone, last gate verdict, and remaining budget sit next to the agent: governance has a read side, not only an enforcement side
11PayMetering + token/inference transparencyCost is attributed per agent, per session, per organization: consumption traces to an entity, not an anonymous key
12RetireLifecycle controlsThe agent is suspended and decommissioned: the same offboarding an employee gets

Author, admit, and mint: bringing an agent into being​

Agents are built, and there is a dedicated shell for building them: a Studio carrying a Mint Agent wizard, a blueprint catalog and editor, skill catalog and editor, function editor, formation composer, billet editor, an Instantiate Crew wizard, a crew monitor, and the agent's own identity and keys. A blueprint is a reusable agent definition that can be cloned and promoted between environments.

Before an agent becomes an actor, admission gates it against policy, with a preview so an operator sees the verdict before committing. Only then is identity minted: Ed25519 principal material and an X.509 certificate from the cell authority, carrying the agent's ID and clearance. Crucially, establishing identity grants no ability to act, a new agent's authority envelope is empty, and capability is granted afterward, explicitly. You can always grant up; you can never un-leak a silent grant.

Brief before work: onboarding for agents​

An unusual step, and one that lands with regulated buyers: an entity is briefed before it is assigned work. Field manuals are authored, sectioned, versioned, and published, bound to spaces and roles; a briefing is resolved for a given entity and an acknowledgement is recorded when it has read it. Onboarding is a surface (a briefing reader and an onboarding checklist) rather than a doctrine document nobody opens.

There is also a deliberate promotion step. Before an agent is trusted with real work, it passes an explicit, auditable promotion, the same as vetting staff before handing over the keys.

Team, run, and coordinate: putting it to work​

A task can be handled by one agent or by a crew. A formation is the shape of a team; a billet is a post in that formation, independent of who fills it; a crew is the formation staffed with agents. Once running, the crew coordinates over a tiered blackboard, posting, claiming, and completing tasks with no manager relay, and a consensus engine reconciles when several agents answer at once. See Orchestration & Agent Teams.

Population itself is governed. Creation, instantiation, and spawning each pass a quota gate; spawn rate is throttled; and quota raises are stepped. The posture is compensation-before-denial, an estate under pressure rebalances before it refuses.

Govern, observe, pay, and retire: the accountability tail​

While it runs, every consequential action routes through the NPG gate, which can refuse at the decision boundary. Alongside enforcement is a read side: an agent's work zone, its last gate verdict, and its remaining budget are visible next to it, in the browser and on the edge device.

Because the agent is an entity, cost is attributed to it, per agent, per session, per organization, so "where did the tokens go?" is answerable against an identity rather than reconstructed from logs. Finally the agent is retired: suspended, then decommissioned, closing the arc the way staff offboarding does.