Multi-agent orchestration
Specialist agents under one orchestrator, with explicit handoffs, retries and a human approval gate before anything irreversible runs.
Core principles
One orchestrator, many specialists
A single planning agent decomposes work and dispatches. Each specialist has narrow tools and one job, so failures stay contained.
Failure is a normal path
Timeouts, bad tool output and refused calls are expected. Retries are bounded, and exhausted retries escalate rather than loop.
Humans hold the gate
Anything that spends money, contacts a person or changes regulated data pauses for explicit approval. The pause is visible, not hidden.
Every run is observable
Each step writes a timestamped trace: who acted, which tool ran, what it cost, and why it stopped.
Agents
Orchestrator
Plans, decomposes, dispatches and reconciles results
Research
Collects and verifies external information
Code
Writes, executes and repairs implementation work
Data
Queries, reconciles and shapes structured records
Outreach
Composes and sends outbound communication
Security
Checks permissions and exposure before release
Compliance
Maps behaviour to the applicable framework
Memory
Holds durable context across runs