Agents

The runtime is the engine. Everything else plugs into it.

#Core

• Runtime — createRuntime, ReAct loop, events, cost + token accounting.
• Tools — how tools are invoked, parallelism, confirmation.
• Skills — prompts + behaviors bundled into reusable personas.
• Delegation — sub-agents, handoffs, shared context.

#Scale

• Durable execution — persist steps, replay, resume.
• Flow control — branching, loops, conditional tool calls, abort signals.
• Runtime guarantees — at-most-once, idempotency contracts, validator insurance.
• Topologies — supervisor · swarm · hierarchical · blackboard.
• Background agents — cron + webhook triggers.
• Speculate — run N candidates, pick best.

#Production

• Human-in-the-loop — approvals, gated tool calls, review queues.
• Self-debug — agents that read their own traces and retry.

#Related

• Package: runtime
• For agents: runtime
• Concepts: runtime
• Memory — persist agent state across runs.
• RAG — retrieval-augmented context injection.
• Reflection — agents that critique and retry their own output.