Team ConnectAI agents
Create and manage the company's AI agents: roles, instructions, the dev team, approvals, company memory and usage limits.
AI agents are robot colleagues who live in the office: they appear on the map, take tasks from the panel or the chat and report what they are doing. They belong to the company, so they stay even if whoever created them leaves. This guide shows how to create them and keep them under control. For everyday use, see AI agents for people.
Who can manage them
Owners, admins and anyone with the Manage AI agents permission create and configure agents. On each agent you can also choose Stewards: people who can configure that agent only.
Open AI agents in the sidebar. The page has the Agents tab (everyone sees it), plus Activity, Memory and Manage.
Create an agent
- On the Manage tab, click New agent.
- Under Start from a template, pick a ready-made template (Reception, Meeting secretary, IT support, Writer or Code reviewer) or click Start blank. Everything stays editable.
- Fill in Name, What it does and Skills (comma-separated, up to 8).
- Under Brain, choose:
- Managed by Team Connect, with one of the company's LLM connections: select the LLM connection, the Model (blank means the connection's default model) and write the Instructions (persona and company rules).
- External: your own program, with the agent's token: your own program reports the agent's state with a token.
- Choose the Role, set Who can ask for tasks and the Robot colour.
- Click Create agent.
Then, on the agent's card, choose its Workstation on the map and use Edit, Pause/Resume, New token or Delete. An external agent's token is shown only once: keep it only in the agent's environment.
Roles
| Role | What it does |
|---|---|
| Assistant | Works on what it is asked, reports steps and asks questions. |
| Code reviewer | Reviews pull requests in the granted repositories and comments on the PR. Never approves or merges. |
| Developer | Changes code on the company workbench, pushes the task branch and opens the pull request. Never merges. |
| QA | Runs a pull request's tests in an isolated environment and reports the result. |
Developer and QA need the managed brain and a configured development workbench.
The dev team
Combine roles to build a team: the Developer opens the pull request and can ask a Code reviewer for a review or a QA agent for tests. The request becomes a task for the other agent, which must have access to the same repository. There is a cap on hand-offs per task, to avoid loops. Merging is always done by a person.
For Developer agents, the Default coding engine field sets which coding engine (Aider, Claude Code or Codex CLI) the agent uses when it hands off a heavy step. It only applies to engines enabled on the workbench and compatible with the agent's LLM provider.
Approvals and tasks
When an agent needs a decision, it asks whoever requested the task. That person answers in the agent's Tasks panel (Approve, Decline or Reply) or in the chat itself. If the workbench is set up for it, the Developer asks for approval before opening each pull request.
Requesters cancel their own tasks. People who manage agents and stewards reorder and cancel any task in the queue. On each task, View transcript shows what was sent to the model and its answers.
Follow the work
The Activity tab shows a board with the columns Queued, In progress, Waiting for an answer, Done and Not done. Filter by Agent and by Finished in the period. Stewards see only the agents they look after.
Company memory
What agents learn is kept by the company, not by the LLM provider: changing models erases nothing.
- Open the Memory tab.
- Filter Status by Proposed to see what agents suggested.
- For each item, click Approve, Edit or Archive. Archived items can be restored or deleted.
- To write a fact by hand, use New memory, choose Applies to (whole company, one repository, one role or one agent) and click Add.
If you prefer, turn on Automatically approve memories proposed by agents. Each task receives only the active memories that apply to it.
Budgets and limits
- Company monthly token limit: set on the LLM providers card. See LLM connections.
- The agent's Monthly token limit: in the agent form. When blank, only the company limit applies.
When the limit is reached, the agent shows as paused with Monthly budget used up and takes no new work until the month turns. Long development tasks also have their own caps on steps, tokens and time.
Privacy
Task content goes through the chosen LLM provider. Transcripts stay in the company database for the retention period you set, and agents' actions are recorded in the audit log.