Skip to content

Capability map

This map is the task-oriented index to the complete documentation set. The curated navigation keeps the common path short; this page makes every implemented capability and design record reachable without knowing its name in advance. For exact signatures, use the generated API reference. For executable code, use the runnable cookbook.

Agent authoring and execution

  • Agent identity — bind a run to the caller it acts for.
  • Activity policies — bound retries, timeouts, and activity behavior.
  • Audit — record security- and operation-relevant decisions.
  • Autonomy — declare how independently an agent may act.
  • Delegation and peer delegation — narrow authority across child and remote work.
  • Dispatch — derive an execution schedule from declared dependencies.
  • Estimation — estimate a run before admitting live spend.
  • Execution models — choose in-process, queued, or durable execution.
  • Parallel fan-out — aggregate concurrent branches without hiding failures.
  • Planning — separate model-authored plans from deterministic execution.
  • Suspension — persist approval waits and resume them safely.
  • Compensation — account for effects that need deterministic reversal.
  • Escalation ladder — choose the smallest agent architecture that fits the task.

Prompts, context, and outputs

Knowledge, memory, and retrieval

Tools, code, and remote boundaries

State, events, and recovery

Testing, evaluation, and observability

Developer workflows

Project records