Skip to content

agents.md

Latest commit

 

History

History
56 lines (43 loc) · 1.84 KB

agents.md

File metadata and controls

56 lines (43 loc) · 1.84 KB
Raw

Agents

AgentRunner is the orchestrator. It owns a SkillsEngine, a tool dispatcher, an event bus, and the conversation state.

Lifecycle

from skillengine import AgentConfig, AgentRunner, SkillsConfig, SkillsEngine

engine = SkillsEngine(SkillsConfig(skill_dirs=["./skills"]))
runner = AgentRunner(
    engine=engine,
    config=AgentConfig(
        model="gpt-4o-mini",
        base_url="https://api.openai.com/v1",
        api_key="sk-...",
    ),
)
message = await runner.chat("Help me with X")

What happens on each turn

  1. System prompt build — names + descriptions of eligible skills (budgeted by skill_description_budget, default 16000 chars).
  2. Adapter call — message history sent to the LLM with the tool catalogue.
  3. Tool execution — ToolDispatcher runs each call returned by the LLM and appends results to the history.
  4. Skill loading — when the LLM calls the skill tool, the full body of the requested skill is injected into the next turn.
  5. Repeat until the LLM produces a final assistant message or max_turns is reached.

Events

EventBus exposes 13 event types:

Category Events
Lifecycle AGENT_START, AGENT_END, SESSION_START, SESSION_END
Turn TURN_START, TURN_END
Tool BEFORE_TOOL_CALL (blockable), AFTER_TOOL_RESULT, TOOL_EXECUTION_UPDATE
Context CONTEXT_TRANSFORM, INPUT, COMPACTION
Model MODEL_CHANGE

A BEFORE_TOOL_CALL handler can return ToolCallEventResult(block=True) to veto a call.

Multi-agent

  • connect_mcp_servers([spec, ...]) — pull tools from remote MCP servers.
  • add_handoffs([handoff, ...]) — register transfer_to_<agent> tools that delegate to other agents (local, remote A2A, or any async callable).

See interop.md for the protocol details.