Skip to content

Harness architecture builder

Choose the workload and the contracts it needs. The builder assigns each responsibility to a neutral component, traces the design back to the dossier requirements, and exports a versioned record you can carry into a new harness repository.

Responsibility map
    Versioned design record

    What the builder decides

    The same six responsibilities always exist: model access, context selection, scheduling, tool dispatch, execution, and independent evidence. Your choices add contracts only when the workload needs them:

    • resumable work adds a checkpoint store;
    • approval or effects add a policy and approval gate;
    • irreversible or replay-sensitive effects add an effect journal and reconciliation;
    • delegation adds a supervisor that owns child budgets, provenance, and cancellation;
    • coding adds an explicit workspace boundary.

    The boxes are responsibilities, not required deployment units. One small process may implement several of them, but their contracts should remain visible so authority, state, and evidence do not collapse into a framework callback.

    Export the same design locally

    python -m examples.harness.design_lab --workload coding --effect-risk irreversible --restart effect_safe --approval risky --delegation
    python -m examples.harness.design_lab --workload research --effect-risk read_only --restart stateless --approval none --format markdown
    pytest tests/test_design_lab.py -q
    

    The JSON and Markdown exports use agentic-arena.harness-design/v1. They contain decisions, components, traced requirements, acceptance experiments, and an explicit evidence boundary. Paste the Markdown into the full design dossier as a starting decision record, then replace assumptions with experiment results.

    This builder produces design guidance. It does not prove that an implementation is secure, durable, or effective.

    Next: Evidence workspace ยท Design dossier