Canonical set

Five reference flows

Five reusable reference flows for common agent-automation contexts. PR review and refund approval are the primary examples because they illustrate both engineer-facing and business-facing evaluator designs across the full task contract. Flows are domains; lifecycle is the path every domain task travels.

Flows are where the task model becomes concrete: each card shows the task boundary, the evaluator, the evidence logged, what earns promotion to more autonomy, what triggers demotion, and what should never become fully autonomous. For the primitive behind every gate, see Evaluators. For the illustrative task declaration rendering, see Framework.

Stacked semantic review

PR review agent

Best for: merge safety
Worker: LLM + tools
Evaluator: Deterministic tests and lint run first, then an LLM judge handles the semantic call the deterministic layer cannot make.
Target: output
Risk: medium / technically reversible, operationally costly
Trajectory: HITL → HOTL

Read the flow

Selective autonomy classifier

Refund approval agent

Best for: selective autonomy
Worker: Rules, classifier, or LLM depending on the implementation.
Evaluator: Learned escalation classifier trained on accumulated human decisions and tuned for escalation recall.
Target: output
Risk: medium / partly reversible
Trajectory: HITL → AUTONOMOUS

Read the flow

Non-generative classification pipeline

Document routing

Best for: low-risk classifier autonomy
Worker: A routing classifier with no generative model in the loop.
Evaluator: Schema validation, out-of-distribution detection, and low-confidence escalation in flight, with downstream reroute monitoring behind the scenes.
Target: input and output
Risk: low / reversible
Trajectory: HITL → AUTONOMOUS

Read the flow

Irreversible action gate

Transaction / trade execution gate

Best for: irreversible action control
Worker: Upstream automation proposes the transaction or trade.
Evaluator: Position limits, price bounds, whitelists, and kill-switch conditions gate the action. Model signals remain advisory only.
Target: output and outcome
Risk: catastrophic / irreversible
Trajectory: HITL

Read the flow

Evaluator maturity pipeline

Support response agent

Best for: evaluator maturity
Worker: The LLM that drafts the response.
Evaluator: LLM judge first, then a classifier distilled from the judge's labels once enough examples accumulate.
Target: output
Risk: low-medium / reversible
Trajectory: HITL → HOTL

Read the flow

What each flow proves

Every reference flow demonstrates a distinct domain, risk profile, and teaching point. None is redundant with another.

Flow proof matrix

Flow	Pattern	Risk / reversibility	What it proves
PR review agent	Stacked semantic review	medium / technically reversible, operationally costly	Cheap deterministic checks can stack under an expensive semantic judge, and the artifact boundary itself is the thing being reviewed.
Refund approval agent	Selective autonomy classifier	medium / partly reversible	Worker and evaluator are different roles, and selective autonomy means the same task can be autonomous for one slice and gated for another.
Document routing	Non-generative classification pipeline	low / reversible	The live evaluator and the promotion evidence are different things, and both are required if the automation is going to stay safe.
Transaction / trade execution gate	Irreversible action gate	catastrophic / irreversible	Risk-adjusted failure cost sets the bar, and irreversible actions stay under deterministic gating rather than learned judgment.
Support response agent	Evaluator maturity pipeline	low-medium / reversible	Eval-the-eval is the point here, because the evaluator is validated, promoted, demoted, and eventually made cheap enough to own the gate.