Two's Complement Speed of Thought
Apr 11, 2026
They dig into running autonomous code-improving agents as Git remotes and the Docker container workflows that keep them isolated. The conversation covers agent teams, incremental commits, and using TODO comments to steer edits. Security, cloud dev environments, vendor reliability, and how agent tooling changes what programming feels like all come up.
AI Snips
Chapters
Transcript
Episode notes
Push Broken Code To An Autonomous Sandbox
- Do push partially finished code into a sandboxed remote that runs autonomous agents to iterate on it.
- Ben uses a Docker Git remote that triggers Claude-P in a post-receive hook to make incremental commits into a working copy for later review.
Treat Agent Work As Draft Commits You Review
- Insight: Treat agent outputs as a draft workflow where you review commits rather than micromanage.
- Ben's pipeline creates working copies inside Docker, lets Claude commit there, and then you pull and review those commits into your main repo.
Run Agents In Sandboxed Containers And Monitor Output
- Do run agents in tightly sandboxed environments and avoid interactive babysitting.
- Ben runs Claude-P in Docker with restricted permissions and watches outputs in tmux for debugging, only reviewing final commits.