Vercel Helps AI Coding Agents Continue After a Crash

AI coding agents can work on a task for minutes or hours. But a crash or a time limit can stop them before they finish. Moving an agent to the cloud does not solve this problem by itself. A cloud computer still needs a reliable record of the work already done.

Vercel introduces durable-harness to help agents continue after these interruptions. It supports tools such as Pi, OpenCode, Codex, and Claude Code. A harness is the software around an agent that manages its tools and session. Durable means the session can survive when the process running it stops. The package adds this support without changing the harness itself.

The package keeps a log, or a saved record, of each step. Before a model or tool call starts, it records the planned action. When the call finishes, it records the result as well. If the process stops, another process can read the log and continue. It uses saved results instead of repeating calls that have already finished.

The article gives an example of an agent adding a pricing page. The agent edits files, opens a pull request, and waits for code checks. If a timeout happens during that wait, the saved record helps it continue. However, unfinished tool calls need different rules because some actions change real things. A lookup can run again, but a refund must not happen twice. The model receives information about an uncertain action so it can check.

Developers can use different places to store the session's records and files. These include local files, supported databases, and Vercel's cloud tools. Some work can use a small virtual workspace, while other work needs a separate sandbox. A sandbox gives the agent its own place to run commands and edit files. The main goal is to continue interrupted work while respecting each tool's retry rules.