Conducting Coding Agents with an Audit-Grade Paper Trail

What it does

Bernstein is a deterministic Python scheduler that farms out a single coding goal to multiple CLI agents—Claude Code, Codex, Gemini CLI, and over 40 others—running each in an isolated git worktree. It merges only the branches that pass tests, while logging every scheduling decision in an HMAC-SHA256 chain and tracking per-file lineage back to the agent, model, and prompt that produced it. The whole coordination loop is plain Python; no LLM decides who runs where.

The interesting bit

The author named the project after Leonard Bernstein because it conducts agents like an orchestra, but the real twist is the compliance-first design: detached JWS-signed agent cards, Ed25519 keys, RFC 2104 audit chains, and an air-gap deployment profile that denies outbound by default. That is a lot of cryptography and operational rigor for a solo open-source tool whose primary job is wrangling CLI wrappers.

Key highlights

• 44 CLI agent adapters, from Claude Code and Codex to Aider and sovereign on-prem endpoints, auto-discovered and mixable in one run.
• Deterministic scheduling with zero LLM in the loop: replay a plan and you get the same task graph, which the author claims cut a $400/month nondeterministic merge problem down to size.
• HMAC-chained audit logs, per-artefact lineage (file writes traced to prompt SHA, model, and cost), and signed agent cards using Ed25519/JWS over canonicalized JSON.
• Air-gap deploy profile that denies outbound traffic by default, plus bearer-token auth and per-session JWT scoping so credentials stay in your environment and out of client repos.
• Optional YAML workflow manifests for declarative DAGs when a single goal string is too coarse.

Caveats

• Solo-maintained and explicitly on-prem only: there is no vendor SLA, support contract, or SaaS wrapper, which the README admits means GitHub issues are your support channel.
• The scheduler is deterministic and walled off from emergent agent collaboration, so research-style “let the agents talk to each other” use cases are explicitly out of scope.
• It is narrowly focused on CLI coding agents; non-coding tasks like research or data analysis pipelines are not supported.

Verdict

Engineering teams running three or more coding agents in parallel, especially in compliance-sensitive or air-gapped environments, should look here. If you just want a single chat pair-programmer or a SaaS solution with a credit-card form, this is not your tool.