Lea and agents
Every project has one lead agent. Minga calls her Lea by default; you can rename her when you create the project. She owns the plan: she breaks goals into milestones and items, gives work to workers, reviews results against each item’s checks, and keeps the brief and the plan up to date.
Persistent agents and workers
Section titled “Persistent agents and workers”| Persistent agents (Lea, role agents) | Workers | |
|---|---|---|
| Lives | As long as the project | For one item |
| Does | Plans, reviews, supervises, talks with you | The work: code, designs, research |
| Has | A chat, a mission and instructions, an autonomy level, a place where it runs | A ref (W-12), a run on a runner, a worktree |
| Reads | The brief, goals and milestones, what changed | The brief, its item’s full spec and checks, the specs above it, inputs passed to it |
Lea’s own turns plan and supervise; they do not change your code. Work on the repo happens in workers. See Workers and runners.
Role agents. You can add more persistent agents to a project, each for an area (for example Release, or Growth), with Hire an agent on Team. They work like Lea within their area, and Lea supervises their items. Adding them is optional.
An agent works in turns. Something wakes it, it reads what changed, acts with Minga’s tools (read the plan, change it, comment, ask, review), and ends the turn. Lea writes in the chat only when she did something, proposes something or needs something; otherwise her turn ends without a message.
By default an agent’s turns load only Minga’s tools. To let Lea call the MCP servers you configured in your CLI (an analytics or issue tracker server, say) while she plans, turn on Use your CLI’s connected tools in its turns for her. Otherwise she puts outside-service work into items for workers.
What wakes Lea
Section titled “What wakes Lea”Wakes arrive in batches: several things that happen close together become one turn, and wakes that are already handled are dropped before the turn starts.
| Wake | When |
|---|---|
| A message or a mention | You write in her chat, or @mention her in a comment |
| Something changed | A person marks an item done, cancels it, sends it back, moves, edits, reassigns or pauses it, answers an escalation card, edits a goal, or approves a review |
| Another agent’s change | Another agent edits something she plans around |
| A new KPI number | Anyone else records a (non-provisional) number for a KPI of an active goal |
| A run ended | A worker on one of her items finished, got blocked or claimed the item is impossible |
| A review is due | An item she owns waits for her review |
| Children closed | Every item under a container closed; she closes it or plans the next phase |
| Triage | Something new landed in triage |
| A card was decided | You answered one of her cards |
| An ask | Another agent asked her something, or answered her |
| Overdue, spec changed | An item went past its due date; a spec changed while a run was reading an older version |
When a person changed something, her instructions say: they did it on purpose; adjust around it, never undo it.
Agents asking agents
Section titled “Agents asking agents”Agents talk on the record. An agent can ask any other agent in the project (ask with to: "@rio"), and an agent’s
@mention in a comment wakes the agent it mentions. Every ask and answer is a comment on the item it is about.
The chain breaker. When agents wake each other four times in a row without a person in between, the fourth wake is held, and the person who looks after the agent gets a card: “Lea and Rio went back and forth 4 times without a person. Last: …”, with Continue and Stop. Any action by a person resets the count. Worker reports, review requests and children-closed wakes do not count, so the normal flow of work never hits it.
Attempts. An item gets at most 5 attempts that changed nothing (no new output, no push) and at most 20 attempts of any kind since a person last acted on it. After that, only a person can continue it.
Pause (in Project settings › Lea, or Agents for other agents) stops an agent: it takes no turn, not even for a chat message, and no new worker starts on the items it owns. A message you send gets “@lea is paused. Your message waits: @lea reads it when resumed.” Her wakes are kept. On Resume the messages and wakes arrive as one batch, introduced with “While you were paused”. Stopping a single run is always possible, at every autonomy level.
Where Lea runs
Section titled “Where Lea runs”Each turn runs either on one of your runners, in the Claude Code or Codex CLI, or in Minga’s cloud with your workspace’s Anthropic or OpenAI API key. See Choose where Lea runs.
How far she goes on her own
Section titled “How far she goes on her own”That is one setting, her autonomy level.