Long-running agent harness

• Work spans external systems and cannot finish in one request.
• You need reliability across crashes, deploys, or network blips.
• Users expect progress tracking and the ability to resume later.

Harnesses make complex agents production-ready. PMs decide checkpoint granularity, retention, and user controls. They affect infra cost and support load: better harnesses mean fewer stuck jobs and clearer incident handling.

Persist state at each step; store tool results and decisions. Add heartbeats, retries with backoff, and idempotent tool calls. In 2026, expose a lightweight console showing state, cost, and approvals; auto-abort after SLA breaches and notify owners.

A data migration agent runs 45-minute jobs with checkpoints every table. After adding the harness, failed jobs auto-resume 90% of the time, reducing manual recovery effort from hours to minutes while keeping cost per run predictable.

• Relying on in-memory agent state, losing everything on failures.
• No idempotency on tools, causing duplicate side effects after retries.
• Lack of user visibility, so people cancel jobs prematurely.