OUTCOME PLAYBOOK
Govern an AI agent
Authenticate an agent, authorize one exact action, execute it durably, and retain evidence without creating competing policy owners.
- 1. Choose
- 2. Install
- 3. Run
- 4. Verify
- 5. Harden
- 6. Operate
#Choose the package boundaries
Required packages establish the local success path. Production and operate packages close durability and evidence boundaries.
Required
1. Orchestrate
@absolutejs/agent-runtimeRuns the model/tool loop and emits typed lifecycle events.
Required2. Authorize
@absolutejs/agencyOwns the single authoritative action policy and decision ledger.
Production3. Execute
@absolutejs/executionMakes consequential effects resumable, idempotent, and reconcilable.
Operate4. Prove
@absolutejs/auditRecords minimized run, action, handoff, and outcome evidence.
#Prerequisites
An application-authenticated actor and tenant boundary.
A model provider configured outside the policy layer.
A stable idempotency key for every consequential action.
#Install the complete surface
BASH
bun add @absolutejs/agent-runtime @absolutejs/agency @absolutejs/execution @absolutejs/auditSmallest useful file tree
TXT
src/
agent.ts # runtime and tool registration
policy.ts # one Agency owner and action policy
operations.ts # Execution store and effect reconciliation
audit.ts # redacted run/action evidence
server.ts # authenticated transport boundary#Run and verify
Own policy once
Create one Agency instance and pass it to every protected tool surface.
Proof of success: A denied test action produces a decision but no effect.
- 1. Choose
- 2. Install
- 3. Run
- 4. Verify
- 5. Harden
- 6. Operate
#Expected results
An unauthorized action is rejected before its effect runs.
An authorized action has one stable operation id and one Agency decision owner.
Audit records correlate the agent run, action, effect, and outcome without storing prompts by default.
#Development to production
DevelopmentMemory Agency and operation storesProductionDurable tenant-fenced stores
Decisions and effects must survive process loss and horizontal scaling.DevelopmentInline effect executionProductionExecution leases with reconciliation
Timeouts cannot safely imply that an external effect did not happen.#Failure decisions
The action is denied.
Check firstConfirm the actor, delegation, resource, and requested capability match the Agency policy input.
ThenInspect policy precedence and denial evidence.
The effect timed out or its result is ambiguous.
Check firstLook up the existing operation by idempotency key before retrying.
ThenResume or reconcile that operation instead of issuing a second effect.