Agents Reference

The agent is built on the native engine (backend/engine/conversation_loop.py::run_conversation) — a supervisor (orchestrator) with two subagents.

orchestrator

The single supervisor, entry point for every message. Decides whether to answer directly or delegate via the native delegate_to_subagent tool. Its own tools: create_artifact, save_memory, get_memory, delete_memory, plus access to RAG tools to answer simple questions about already-indexed content without delegating.

coder

Specialist in filesystem, terminal, and git. Receives explicit instructions from the orchestrator (doesn’t decide on its own what to do — executes what was delegated). Tools: file_read, file_edit, file_write, grep, list_dir, terminal, all 14 git operations, plus memory and RAG access.

search

Specialist in real-time web search and RAG. Tools: web_search, web_fetch, vector_search, embedding, ingest_docs, plus memory. There’s no separate third subagent dedicated only to RAG — that responsibility belongs to search.

Middleware

HITL (HumanInTheLoopMiddleware) — pauses before destructive tool calls, with behavior configurable by permission mode.
Context — every invocation carries user_id, workspace_id, and permission_mode via a typed context_schema.

Checkpointer

AsyncSqliteSaver (lite mode) or a Postgres equivalent (complete mode) — persists graph state per thread, letting you resume a conversation exactly where it left off.