Task profile

Task Profile

Overview

Run the second stage of the framework. Assign a delegation mode to each process step rather than to a whole workflow, and keep cognitive delegability separate from the technical execution boundary.

When to use

• When a non-trivial workflow mixes exact operations, interpretation, and judgment.
• When deciding how much autonomy an agent receives for each step.
• After workflow-readiness returns ready for task profiling, or within the bounded evidence-gathering scope of pilot/assistive only.

When NOT to use: a one-step read-only request; a task already profiled against unchanged premises this session; or an unbounded recurring workflow that has not passed the readiness screen.

Process

1. State the goal and effective point. Define when each output becomes operationally consequential. Stop for clarification if the goal is ambiguous.
2. Decompose into process steps. Use 3–7 steps at the level where one step can have one route, one verification criterion, and—if agentic—one mode.
3. Route deterministic operations first. Send calculations, threshold comparisons, extraction, conversion, and other fully determined operations to a script, query, or exact tool. An agent may orchestrate the tool but must not replace its specified logic with model inference. Mark the operation not agentic; if its output can take effect, profile orchestration and release as separate steps and apply both gates to them.
4. Rate the remaining agentic step:
   • Structure — can inputs, dependencies, and procedure be specified ex ante?
   • Verifiability — is a shared criterion available at the point the output would take effect? Name it; plausibility is not a criterion.
   • Judgment share — must the human frame the problem, set or weight criteria, or choose among legitimate options?
5. Assign a provisional mode from those three task dimensions:
   • autonomous candidate: structured, verifiable at effect, low judgment;
   • assistive: bounded analysis components are verifiable, while a human must independently weigh a non-binding recommendation and decide;
   • human-final: problem framing, criteria weighting, and option selection must remain human because no shared criterion can settle the choice. The agent may gather evidence and simulate consequences, but may not decide, approve, or implement within that decision step. Profile the chosen implementation separately after the human decides.
6. Apply the per-step architecture gate. Recheck state/constraint retention, error containment, checkpoint/restart, logging, auditability, and recovery for the concrete implementation; the coarse readiness screen does not replace this gate. A failed or unknown gate forbids autonomous. It permits controlled only if an effective pre-effect checkpoint contains the failure and a human can apply the criterion there; otherwise constrain the agent to non-effecting assistance.
7. Apply the governance gate. Check error tolerance, affected parties, reversibility, consequence, and applicable rules. High consequence alone changes an autonomous candidate to controlled; it does not create human-final. Controlled is valid only when a human can compare the output with a named criterion before it takes effect. Without such a criterion, choose assistive or human-final according to who frames and decides.
8. Set the execution boundary from blast radius using references/delegation-modes.md. Blast radius determines permissions, sandboxing, release mechanics, and rollback—never the cognitive mode itself.
9. Checkpoint. Fill assets/task-profile.md, distinguish provisional from final mode, and obtain human confirmation before acting.

Output

For each step, record route, criterion at effect, both gate results, blast radius, and technical boundary. For each agentic step also record the three task dimensions, provisional mode, and one canonical final mode. Mark exact-tool operations not agentic rather than inventing a fifth delegation mode. Name every judgment node.

Rationalizations

Excuse	Rebuttal
"The workflow is routine, so one mode is enough."	Mixed profiles are normal; classify each process step.
"A reviewer makes it controlled."	Controlled requires a criterion the reviewer can actually apply before effect.
"High consequence means the human must do everything."	Consequence adds a release gate to verifiable work; it does not manufacture judgment.
"The blast radius is six, so the mode is human-final."	Blast radius sets the execution boundary, not the source of judgment.
"The architecture is probably reliable."	Unknown architecture evidence excludes autonomous delegation.

Red flags

• controlled with no named criterion available at the gate.
• autonomous despite an applicable readiness cap or a failed architecture gate.
• human-final used merely because consequences are high.
• Model inference performing an exact calculation or threshold rule.
• Blast radius used as a hidden fourth task dimension.

Verification

The profile is complete when every step has a route, named criterion or explicit absence, both gate results, and a separately stated execution boundary; every agentic step has three task ratings and one canonical final mode; every exact operation is marked not agentic; required readiness caps are honored; all judgment nodes are named; and the human confirmed the profile before action.