sub-agent-orchestrator

Define a YAML workflow specifying agent roles, prompts, tools, and steps, then run it through the orchestrator. Steps execute in topological order with each agent's output stored and passed to dependent steps automatically.

Sub-agent orchestration coordinates heterogeneous pipelines where agents have different roles and outputs feed subsequent inputs. Agent swarms deploy homogeneous agents for parallel data processing where items are independent with no task-level dependencies.

Use a parallel step type listing multiple agents with output keys, set a wait policy (all, any, or N), then add a sequential aggregation step that receives the collected outputs. The lead scoring example shows three scorers fanning out into one aggregator.

Failures trigger the agent's retry policy with configurable max attempts and backoff. If retries are exhausted, the on_failure policy applies: skip stores null and continues, abort stops the workflow, and fallback deploys a backup agent with the same input.

Yes, each agent supports validation with JSON schemas and plain-English rules such as field presence, array length, and value ranges. Failed validation is treated as a failure and triggers the retry or fallback policy.

Avoid it for single-step tasks or homogeneous bulk operations like mass code changes or independent data processing, where Agent Army or Agent Swarm patterns fit better. Circular dependencies are also rejected at parse time.