What problem does it solve?
Many AI products fail not because the model is weak but because the harness around it is fragile: unclear tool boundaries, missing permission policies, brittle workflow state, stale context, weak evaluation, and poor operator visibility. This Skill turns those vague operational and safety gaps into concrete primitives, boundaries, phased plans, and verifiable acceptance checks so agents behave predictably and safely in production.
Core Features & Use Cases
- Harness design & classification: Choose the right harness shape (chat assistant, code agent, workflow orchestrator, embedded feature, or hybrid) and state the smallest viable subsystem set.
- Permissions & tool governance: Define capability registries, permission tiers, approval gates, and safe execution boundaries to prevent unauthorized side effects.
- Durability & state management: Specify workflow state machines, idempotency, retries, and resumability for long-running or approval-driven tasks.
- Context, memory & evaluation: Enforce context budget discipline, provenance-aware retrieval, memory rules, and a lean evaluation system with golden tasks and regression checks.
- Operational UX & observability: Design user-facing status, logs, health checks, and cost visibility so operators can diagnose and control runs.
- Use case example: Turn a brittle coding assistant that accidentally edits files into a solo-maintainable harness with approval gates for writes, tool filtering by context, and replayable evals.
Quick Start
Design a lean, solo-dev harness for a code agent that requires approval for writes, durable workflow state, and a capability registry.
