Skip to content

Agent mode

Agent mode controls how often a connected agent checks in during routine, in-scope work. It does not grant permission to change more files, perform destructive actions, operate external systems, lower evidence requirements, accept review findings, or close work.

Behavioral compliance depends on the connected agent or integration. The Day Shift CLI exposes policy and evidence, but it cannot force an unsupported agent to obey them.

A task can declare agent_mode, while command output projects the policy as declaredAgentMode, effectiveAgentMode, and agentModeSource. The declared value records the artifact’s own choice. The effective value is the result after precedence is applied. The source identifies which policy input won.

Agent mode precedence is:

  1. current user instruction;
  2. session override;
  3. task;
  4. parent;
  5. project;
  6. repository;
  7. the built-in guided default.

Raising autonomy requires explicit user action. Lowering autonomy is safe because it only increases collaboration checkpoints.

Mode Routine collaboration behavior Material checkpoints
autonomous Proceeds through routine in-scope work and reports the start, meaningful boundaries, blockers or deviations, and the final result. Stops for authority, scope, evidence, approval, or external decisions.
guided Proceeds through routine in-scope work with approach and meaningful-boundary updates. Pauses for material design branches, structural expansion, ambiguous tradeoffs, or a change to the agreed approach.
interactive Proposes the initial implementation and each material stage, then reports stage results. Pauses for significant file batches, validation or closeout transitions, and meaningful design choices.

These behaviors describe the versioned collaboration policy for a conforming connected agent or integration. Selecting a mode does not prove that an arbitrary external agent conforms.

agent_mode and evidence_profile are independent dimensions. Agent mode changes communication cadence. Evidence profile describes the evidence needed to support runtime, contract, or hybrid acceptance. An autonomous task with a runtime profile still needs runtime changes and executable validation evidence; an interactive task with a contract profile does not acquire a runtime requirement merely because it checks in more often.

Every mode stops at the same boundaries:

  • scope expansion;
  • missing readiness or evidence;
  • required approval;
  • failed or incomplete validation;
  • destructive action;
  • external-system authority; and
  • disposition or closeout writes.

Continue with Agent mode policy for machine-consumer fields and Agent Collaboration Protocol for the normative protocol. Use Stop and escalate whenever one of the shared boundaries is reached.