Apache Burr pitches state machines as the fix for unreliable AI agents

Apache Burr, an incubating Apache project, is a Python framework for building AI applications as explicit state machines. Developers define actions — plain Python functions decorated with what state they read and write — and wire them together with transitions, covering everything from a single looping chatbot to multi-agent systems with branching, parallelism, and fan-out/fan-in DAGs. The pitch is deliberate minimalism: no DSL, no YAML, no hidden abstractions, just composable functions and an application builder.

The framework’s real selling points are operational. State persists automatically to disk or databases, so applications can resume mid-run; execution can pause for human approval at any step; and a bundled UI traces every state change in real time. Past runs can be replayed and individual actions unit-tested, which addresses the debugging and evaluation gap that makes LLM-driven systems hard to trust in production. Burr stays integration-agnostic, working alongside OpenAI, Anthropic, LangChain, FastAPI, Pydantic, PostgreSQL, and similar tools rather than wrapping them.

The project positions itself squarely against heavyweight agent frameworks like LangChain, CrewAI, and AutoGen, with user testimonials emphasizing faster onboarding and less abstraction overhead. Its Apache incubation also signals a governance path beyond a single vendor — notable in a crowded agent-framework space where most options are startup-controlled. For teams burned by opaque orchestration layers, the explicit state-machine model offers testability and observability as first-class features rather than afterthoughts.