A company brain that navigates live sources instead of embedding them

What it does

Scout is a single Python agent that talks to your company’s scattered sources—Slack, Google Drive, web, wikis, CRMs, and MCP servers—through a uniform layer of natural-language query and update tools. Rather than bulk-ingesting everything into embeddings, it navigates sources live when you ask, then persists what it learns into its own auto-managed wiki and CRM. It is built on the Agno framework and designed to run as a local or deployed service with an optional web UI via AgentOS.

The interesting bit

The architecture hides each source’s quirks behind a sub-agent. Scout sees a clean query_slack tool; behind it, a dedicated agent handles cursor pagination, user lookups, and thread semantics so the main agent’s context stays focused on your question instead of Slack API trivia. It applies the same “navigation over search” insight coding agents use—ls, grep, open file—to company data.

Key highlights

• Treats live sources as a navigable filesystem rather than a vector dump, querying on demand.
• Auto-maintains a wiki and CRM: learns facts like contacts or shared papers and schemas them on demand via agentic SQL.
• Context providers expose only two natural-language tools per source, reducing context pollution and scope overlap.
• Ships with eval tiers: code-level wiring checks, behavioral cases, and LLM-scored judges.
• Production deployments enforce RBAC-signed JWTs by default; the filesystem wiki is ephemeral unless you switch to a Git backend.

Caveats

• Several promised providers (GitHub, Gmail, Calendar) are built but not yet landed in the main branch, pending testing with real tokens.
• The default filesystem wiki resets on every container restart, so production requires manual Git-backend configuration to persist knowledge.
• It is tightly coupled to the Agno ecosystem and AgentOS control plane for auth, UI, traces, and metrics.

Verdict

Worth a look if you are building internal AI tools and tired of ingest-everything vector pipelines that chunk, embed, and pray. Skip it if you need a finished, provider-agnostic product with zero framework coupling.