VercelEngineering PostOfficial Source

Making every harness durable

Vercel’s open-source durable-harness logs model and tool steps so cloud coding agents can resume after crashes without rerunning completed calls.

Vercel · Oct 19, 2026

Open Source Open Markdown Open JSON

Source Summary

Vercel’s open-source **durable-harness** adds crash recovery to Pi, OpenCode, Codex, Claude Code, and custom harnesses. It records each model or tool-call intent and outcome, then backs the turn with a queue message.

Practical Implication

For long-running coding agents, persist execution state outside the process. After a crash or timeout, a new function reads the log, restores harness session records, and resumes; **completed calls do not run again**, avoiding repeated model charges.

Agent-Ready Context

Vercel’s open-source **durable-harness** adds crash recovery to Pi, OpenCode, Codex, Claude Code, and custom harnesses. It records each model or tool-call intent and outcome, then backs the turn with a queue message.

For long-running coding agents, persist execution state outside the process. After a crash or timeout, a new function reads the log, restores harness session records, and resumes; **completed calls do not run again**, avoiding repeated model charges.

Tool recovery is not universally automatic. Interrupted tools marked **replay: 'safe'** run again, while other tools return as possibly executed and leave the model to decide whether to inspect before continuing.

Connected Context · Feed7 Judgment

This turns the prior durability guidance into a reusable implementation across several coding-agent harnesses. Its intent/outcome log and queue-backed turns address crash recovery and duplicate model charges, while replay labels expose the remaining hard boundary: an interrupted external tool may have executed. Recovery therefore still requires idempotent tools or post-failure inspection, not blind replay.

AI Agents Are Just Distributed Systems Now — Salman Munaf, TikTokDurable Harness implements the persisted-step and recovery side of the distributed-systems guidance, while its possibly-executed tool state confirms that timeout ambiguity still requires idempotency or reconciliation.Every step you take, every call you make: the reliable agent stack — Giselle van Dongen, RestateRestate frames durable state, retries, and recovery as infrastructure concerns; Durable Harness supplies a concrete queue-and-log design for applying that pattern to existing coding harnesses.What we’ve learned building cloud agentsCursor’s Temporal rewrite provides reinforcing operational evidence that durable orchestration matters for cloud-agent reliability, while Durable Harness makes a related recovery pattern portable across multiple harnesses.Agent Frameworks Considered Harmful — Rémi Louf, .txtThe append-only intents and outcomes, queue-backed turns, and replay handling closely instantiate the candidate’s call for durable events, causal traces, and replay-oriented runtime primitives.

Context Map

agentcoding#harness-engineering#agent-reliability#cloud-agents

Uncertainty

Tool recovery is not universally automatic. Interrupted tools marked **replay: 'safe'** run again, while other tools return as possibly executed and leave the model to decide whether to inspect before continuing.