Use Cases
Internal copilot
Build an internal assistant over company data with RAG, permissions, observability, and a path to production.
An internal copilot usually starts as “chat over docs” and quickly becomes more demanding:
- company knowledge and retrieval
- multiple data sources
- user or team-specific permissions
- auditability and production controls
Typical stack
npm install @agentskit/react @agentskit/adapters @agentskit/runtime @agentskit/memory @agentskit/rag @agentskit/observability @agentskit/toolsRecommended package mix
| Layer | Package | Why it matters |
|---|---|---|
| UI | @agentskit/react | Fast way to ship an internal assistant surface |
| Provider | @agentskit/adapters | Model flexibility as cost and privacy needs change |
| Runtime | @agentskit/runtime | Lets the assistant retrieve, act, and chain tool calls |
| Memory | @agentskit/memory | Preserve ongoing user context and preferences |
| RAG | @agentskit/rag | Ingest, retrieve, rerank, and connect knowledge sources |
| Tools | @agentskit/tools | Reach internal systems through integrations or MCP |
| Ops | @agentskit/observability | Critical for trust, debugging, and access-sensitive workflows |
What the architecture usually looks like
- A user asks a question in a chat UI.
- The runtime combines short-term context, memory, and retrieved documents.
- Tools fetch additional data from internal systems when needed.
- The assistant returns an answer, citation set, or action result.
- Observability and security layers provide the controls needed for rollout.
Good defaults
- Start with retrieval-only before enabling mutating actions.
- Keep retrieval quality visible with traces and sampled evaluations.
- Model access and tool access should be treated separately.
- Add prompt injection and PII protections before broader rollout.
Best follow-up guides
When AgentsKit is especially strong here
Internal copilots are where the ecosystem framing matters most. AgentsKit gives you one set of contracts across UI, retrieval, runtime, tools, and production controls, which helps a lot once the “simple chat over docs” prototype starts growing into a real internal platform.