The twelve factors

Define the outcome, execute within authority, check the exact output, then use what you learned to improve the next cycle. Operating principles, not twelve mandatory steps.

Tier I–III

Brief the Agent

Intent, sources and a bounded job.

IShip Packets, Not TranscriptsBrief the agent with accepted intent, authority, relevant sources and required checks.judgmentdurable context IIGive Every Claim One OwnerKeep intent, records, work status and check evidence with their authoritative source owners.durable contextloop closure IIIOne Agent, One JobGive one agent one bounded outcome; add workers only for useful independent scopes.judgment

Tier IV–VI

Lay the Rails

Authority, methods and write ownership.

IVCap the Blast RadiusBound access, actions and spending with limits the runtime or source system can enforce.judgment VMap the Terrain, Then Pave ItInspect current sources, then encode supported methods in skills, templates and checks.judgmentdurable context VIOne Lane, One WriterGive each writable resource one writer or a working concurrency control.judgment

Tier VII–IX

Gate the Work

Checked outputs, preserved behavior and memory.

VIINo Verdict, Not DoneCheck the exact output against intent; use fresh independent judgment where risk calls for it.judgment VIIIRatchet the BaselinePreserve accepted behavior with proportionate checks, approved examples or source-backed instructions.judgmentloop closure IXClose the Learning LoopKeep linked work history and memory; curate supported findings for later work.durable contextloop closure

Tier X–XII

Govern the Loop

Caller control, stopping boundaries and feedback.

XOwn the Verdict, Rent the HarnessKeep acceptance and evidence under caller control, whichever runtime, orchestrator or factory runs the work.judgment XIStop the LineRepair known defects within scope; stop at missing authority, unresolved intent or spent limits.judgmentloop closure XIIPrice the ProofMeasure cost per accepted output, include failures and repairs, and feed findings into the next brief.loop closure

The evidence

How well research backs each factor.

Each essay cites its sources and says how strong they are. Some rules have been tested directly. Others are practice the research only points toward.

Measured

Studies test the claim behind the rule directly.

Supported

Studies support the reasoning. The claim behind the rule has not been tested head-on.

Practice

Engineering practice. Agent research supports it only indirectly.

One example, twelve essays

Every essay follows the same job.

Draft a monthly operations report from an approved ledger. Reconcile totals, cite claims, exclude personal details and get operator approval before publishing.

The same terms apply when the output is an investigation, a document, an operational action or a code change.

Intent: The caller’s accepted outcome, examples, scope and authority.
Work item: A bounded assignment with status and dependencies in the existing tracker.
Output: The exact candidate result: a report, finding, document, action or code change.
Checks and judgment: Checks establish facts; independent judgment evaluates meaning where risk requires it.
Acceptance and memory: The caller’s policy decides acceptance; linked records and supported findings inform later work.

See the factors at work.

AgentOps is the reference implementation. Install it and run one read-only Research task in a repository you already work on.

Install AgentOps