Chock launches sandbox-first AI coding harness

Chock enters a landscape where coding agents already run commands, write files, and reach the network, often next to production credentials on developer machines. The harness's pitch is that most other harnesses ask the agent to be careful, while Chock puts what an agent may do outside the agent's control. Its policy file, chock.zon, sits in the project and can be read but not edited by the agent; if an action needs approval and nobody answers, it is refused.

The comparison table is based on reading source code, not documentation. It says ten harnesses were read, and the three that sandbox at all mount the whole filesystem into the sandbox, and none keeps a tamper-evident log. Chock's answer is an append-only, hash-chained log where each event carries the hash of the one before it, plus an organisation policy that projects can narrow but not widen.

The limits are explicit. Attribution is worth what rollout coverage is, sealing depends on key custody, and there is no remote attestation or hardware root of trust. Whether this changes how security teams admit agent sessions likely hinges on whether the remaining roadmap links — a signature on the commit and a forge attesting the push — ship, and on whether the log's tamper evidence holds up in an audit.