Sandboxed coding agents

Some questions can only be answered by reading the code. Why does this customer see this error? Has this bug been fixed already? So I made Claude Code and Codex a step you can drop into any pipeline, just like an email or a Slack message. It gets a fresh, throwaway machine with the right repositories and docs, does the investigation, and hands back an answer the rest of the pipeline can use.

Safe by design

Each run happens in its own isolated sandbox that is destroyed afterwards. It can only reach the model, our own API and the code it was given, and its access expires when the run ends. That makes it reasonable to let an AI run commands on real code.

Any model

The step works with both Claude Code and Codex, each with its own models, so a team can use whichever fits them.

Answers, not essays

A pipeline needs data from the agent, not a chat message. The agent hands back a structured result, and if it is incomplete or wrong it is told why and fixes it, so the next step can rely on it. Every run's full transcript can be read afterwards.

The same machinery runs Chip and the sandboxed apps.