Harness Intelligence Wiki
HarnessExecution Modes

Manual HITL Flow

Operator-led harness execution through skill-by-skill boundaries

Manual HITL flow keeps the operator at the workflow boundaries.

It can use the same underlying skills as an AFK run, but the human decides when the next skill has enough evidence to start. The trust model comes from the invoked skills themselves: each one defines what it produces, when it stops, and what the operator should inspect before continuing.

Skill-Level Stops

Manual HITL should not become a separate checklist layered over the harness. The review moments come from the skill being called.

create-spec compiles closed decisions without interviewing or adding an approval stop. The operator may inspect the resulting agent-ready SPEC.md before calling write-backlog. create-plan returns an execution handoff. implement-spec runs worker waves and produces validation evidence plus a manual review checklist when behavior has a user-verifiable surface. Debugging and debt resolution stop when the evidence requires a human decision.

The operating rule is simple: call one skill, inspect the artifact it was designed to produce, then decide whether the next skill has enough evidence to run.

Manual Chain

optional human-selected business-finder | functional-finder
  -> or direct requirements-phase for bounded requirements
  -> requirements-grill
  -> create-spec
  -> write-backlog (derive Epic / Story / Task identities)
  -> create-plan
  -> implement-spec
  -> review-phase
  -> optional debugging-phase
  -> docs-ingest-phase

Finder is optional intake. It may capture a Fog and project Product Areas and Initiatives through Business Finder, or that Business structure plus Epics through Functional Finder. Requirements Phase alone owns Requirements Grill, Create Spec, and delivery-depth Write Backlog projection. Specifications identify accepted outcomes as OUT-###; Write Backlog derives actual Epic, Story, and Task identities after compilation.

The operator may stop between any two steps. Stopping is part of the model when the next step would require guessing.

Use Manual Flow When

Use manual flow when intent is still forming, requirements have unresolved branches, product judgment matters, runtime behavior depends on a human-observed repro, or a debt item needs inspection before it can be accepted.

Done Signal

A manual run is ready to close when the artifact trail is inspectable enough for the operator to continue, close, or hand the work to another mode with confidence.

That is the boundary with Goal Mode And AFK: manual flow keeps the operator present between skill boundaries, while AFK delegates those boundaries to a bounded goal.

On this page