For the CTO
The Orchestration Engine is the BEAM-native runtime for governed agent execution. It solves:- How to run task agents and orchestrators durably
- How to govern decomposition, execution, review, and reflection
- How to supervise sessions and intervene safely
- How to preserve provenance, audit trails, and promotion history
How It Works
- Define work as Nodes in a Graph (problems to solve, features to build, research to do)
- Decompose large problems into smaller Tasks through orchestrator agents
- Assign Tasks to agent sessions with explicit authority, tools, and timeouts
- Execute through governed harnesses (Claude Code, Codex, Gemini, or any adapter)
- Review outputs through mandatory independent review (multi-model GAN or human)
- Reflect on what went well and what the system should learn
- Promote reviewed artifacts through workspace claims and governed mutation
Key Properties
- Declarative API — spec/status pattern, reconciliation-based (like K8s but for agents)
- Model-agnostic — swap Claude for Gemini for a local model; governance doesn’t change
- Policy-based admission — no work starts without passing the governance pipeline
- Monotonically narrowing authority — children can never exceed parent’s grants
- Producer can’t certify own output — structural provenance separation, not just different endpoints
- Budget through timeouts and retry limits — not token counting (that’s billing, not governance)