scout: Instructions file for Codex

AGENTS.md

Project Overview

Scout is an enterprise context agent — a single agno.Agent with N ContextProviders. Ships with WebContextProvider, WorkspaceContextProvider (local files via agno.tools.Workspace), DatabaseContextProvider (the CRM — user's contacts/projects/notes/follow-ups), two WikiContextProviders (a writable knowledge wiki and a read-only voice wiki), SlackContextProvider, GDriveContextProvider, and MCPContextProvider (any MCP server → one query_mcp_<slug> tool on Scout). GitHub, Gmail, and Calendar land in the next release (were built and verified on the feat/slack-interface branch; dropped from the ship slice until we can test end-to-end with real tokens).

Architecture

Scout (single Agent — one LLM hop per turn)
  tools = <query_|update_ tools from every registered ContextProvider> + list_contexts

Every source is a ContextProvider. The database is a provider too: DatabaseContextProvider exposes query_crm (reads) + update_crm (writes), each backed by a dedicated sub-agent so the read path never sees the write engine.

ContextProvider

agno.context.provider defines the base. Every external source subclasses ContextProvider and implements:

• query(question) -> Answer / aquery(question) -> Answer — natural-language read
• status() -> Status / astatus() -> Status — is the source reachable?

Providers that support writes override aupdate(instruction) -> Answer (and optionally update). The base raises NotImplementedError; _update_tool() translates that into a readable " is read-only" error so calling agents see a clean failure. Today only DatabaseContextProvider overrides it.

mode controls how the provider surfaces itself to the calling agent:

Mode	Exposure
default	The provider's recommended exposure. Each subclass decides.
agent	One query_<id> tool wrapping a sub-agent.
tools	The underlying tools directly.

model swaps the model used by the internal sub-agent (when one is built). instructions() returns mode-aware usage guidance for the calling agent.