Observe mode first

DeepSeek just published their design for an always-on coding agent on a team repo. The resource it budgets is your team's attention, not GPU time. Unasked work starts as a log entry, not a PR.

This matters if you put coding agents on shared repos. dsh-superagent (MIT, Oct 31) is a research prototype on DeepSeek Harness. Three days in: 14,121 stars, 1,504 forks. One repo, one agent: people on an allow list ask for work, each job runs in its own session and directory, and the answer comes back as a pull request.

The part worth copying is the authorization model. Proactive work defaults to observe mode: the agent assesses a candidate, records it as "would accept", and does nothing. A person reviews the record, then switches to auto. Permissions live in the host: messages, comments and quoted documents are inputs to a decision and cannot change them. Auto merge only on its own PR, approvals on the head commit, fewer than 200 changed lines. Replies cap at 100 words.

The decision this changes: stop debating whether a team agent is safe to enable. Enable it in observe mode and let the recorded decisions argue for wider scope.

The trap: none of this is isolation. Workers run with full access as the operator's OS user. The docs say it plainly: defaults, not a security boundary. Sandboxing stays your job.

Before you give an agent write access to a team repo, ask what it would have done last week without asking. Observe mode is how you get that answer.