תזמור הצהרתי של סוכנים

AI agent orchestration is the practice of coordinating multiple AI coding agents to work on a software project in parallel. Instead of running one agent at a time, an orchestrator decomposes goals into tasks, assigns them to agents with appropriate models and roles, manages isolation via git worktrees, and verifies results before merging.

Bernstein is one of the open-source options for orchestrating AI coding agents. It supports 37 CLI agents, uses deterministic Python scheduling (no LLM tokens on coordination), provides built-in verification via a janitor system, and has zero vendor lock-in.

Unlike CrewAI and Microsoft Agent Framework (AG2, the successor to AutoGen — AutoGen entered maintenance mode in April 2026), Bernstein uses deterministic Python code for all orchestration decisions instead of LLM-driven scheduling. Bernstein agents are short-lived (minutes per task, not hours), support any CLI agent (not just API-based models), and include built-in janitor verification.

Yes. Bernstein supports 37 CLI agents including Claude Code, Codex CLI, Gemini CLI, Qwen, Aider, Amp, and more. You can mix models in a single run — for example, Claude Opus for architecture tasks, Codex for test generation, and Gemini for documentation — with automatic model routing based on task complexity.

Yes. Bernstein has native adapters for Claude Code (Opus 4, Sonnet 4, Haiku). Claude Code is a first-class citizen in Bernstein, with full support for model selection, effort routing, and all orchestration features including git worktree isolation and janitor verification.

Yes. Bernstein is fully open-source under the Apache 2.0 license. Install via pipx install bernstein or pip install bernstein. The source code, documentation, and issue tracker are on GitHub at github.com/chernistry/bernstein.

Bernstein uses a janitor verification system that checks concrete signals — tests passing, files existing, no regressions — before accepting any agent output. Failed tasks are automatically retried with a fresh agent and clean context. A circuit breaker halts execution if repeated failures indicate a systemic issue.

Deterministic scheduling means the orchestrator uses pure Python code, not LLM calls, to decide task assignment, priority, and agent selection. This eliminates nondeterminism, reduces cost (zero tokens spent on coordination), and makes the system predictable and debuggable. Only the initial goal decomposition uses an LLM.