Stages are good at one thing each. Something has to decide which stages run, in what order, with what inputs, and what to do when one gets stuck. That is the orchestrator. It reads each stage’s declared metadata to decide:
- Inputs and outputs (Compartment 1): what each stage consumes and produces.
- Purpose: what transformation or validation the stage performs.
- Orchestrator hints: sequencing constraints, conditional triggers, escalation policies.
a. Goal ownership
Owns the end-to-end objective. It never hands goal-tracking to a single stage; it checks the combined output still matches the original intent.
b. Workflow composition
Builds a workflow for this job. Greenfield, bug fix and refactor get different sequences. The plan is mutable and adapts as it runs.
c. Routing, observability, control
Feeds each stage the right inputs, tracks state, enforces halting conditions, handles escalations (pause, ask a human, resume) and keeps a full audit trail.
d. Abstraction boundary
Treats each stage as a black box. It only checks the declared post-conditions were met, so stages can evolve independently.
e. Cross-stage invariants
Some rules span everything: naming conventions, security policies, documentation standards. The orchestrator checks these at checkpoints.
Think of the orchestrator as a general contractor: it does not lay bricks, but it decides who works when, checks each trade's work against the contract, and keeps the paper trail.
Want the original wording? Read Section 4 of the specification.