Building agents
An agent is a working setup made shareable: instructions, model, knowledge, and tools, versioned like software.
Anatomy of an agent
| Part | What it does |
|---|---|
| Instructions | The system prompt: what the agent is, how it answers, what it must not do. |
| Model preference | Which provider and model it favors — or let smart routing pick the cheapest model that clears your quality floor. |
| Knowledge bases | Attached document collections it grounds answers in, with citations enforced. |
| Tools | Web search, MCP-backed connectors (GitHub, Jira, Notion, databases…), code execution. |
| Personality | Tone and formatting defaults, so five people deploying it get one consistent voice. |
Versioning and rollback
Every publish creates a version. Editing a live agent never mutates what users are talking to — the new version goes live when you publish, and rolling back to any prior version is one click. Treat prompt changes the way you treat deploys, because that's what they are.
Visibility and sharing
- Private — yours alone, the default.
- Org-shared — visible to your organization's workspace, governed by roles.
- Marketplace-published — listed publicly after review, free or paid. See the marketplace for how publishing and payouts work.
Deployment channels
The same agent definition runs everywhere Akili does:
- In the workspace — the default chat surface.
- As a Slack bot or Microsoft Teams bot (Team plans include five deployments of each).
- Embedded on your own website via the webchat widget.
- Through the API with a generated API key (Pro and up).
Deployed agents get an inbox — a review surface for the conversations your agent is having, so you can see where it shines and where its instructions need work.
Scheduled workflows
Agents can also run on a schedule as multi-step workflows: pull data, generate a report, export it, notify a channel. Useful for the recurring work nobody wants to remember.