Structured Output
Overview
Capability Discovery
For discovery architecture and integration points, see Capability Discovery — Token-Efficient Tool Context.
Emergent Capabilities
Runtime tool forging for AI agents: AgentOS lets agents generate, test, judge-approve, and register new tools mid-decision, run in an in-process node:vm context or a QuickJS WebAssembly instance per call. Multi-agent spawn_specialist included.
Structured Reply
A turn on the full runtime whose reply must match a schema. The request names the schema; the runtime instructs the model, parses and checks the reply, asks again inside the turn when it does not match, and hands the caller the parsed value on the final chunk.
Self-Improving Agents
AgentOS has four opt-in tools that let a running agent change its own behaviour within configured limits: adaptpersonality shifts a HEXACO trait, manageskills turns skills on and off for the session, createworkflow chains tools into a reusable sequence, and selfevaluate scores a response and adjusts runtime parameters. They run in the AgentOS runtime (AgentOS.create() and processRequest()); agent() and agency() do not register them.
Agent Config Export & Import
Export an agent's or an agency's configuration as an object, JSON or YAML
Tool Calling & Lazy Loading
This guide explains how AgentOS wires extensions (tools) and skills (prompt modules) into a tool-calling agent loop, and what “lazy loading” means in practice.