Agents and helpers

The short version

Most jobs are one Mori, thinking and using tools in one thread. Some go faster or better split up, or with a second pair of eyes. Then Mori starts agents:

• helpers take one independent piece each and report back;
• a crew is a Captain who plans and checks, leading members with roles;
• a lane is a second request you sent while she worked, running beside the first;
• your own agents are roles you write once and reuse;
• peers are your other Mori bots, and A2A agents are agents outside Mori.

Every one of them is a model run of its own, shown as a row in the turn's box with its model, tokens, time and a Stop.

When helpers start

Mori decides when a helper helps, and a box appears only when one really starts:

• never when you said no helpers, and never for small asks;
• at Instant and Quick depth, only for serious research or heavy jobs;
• at Balanced and above, for serious research or a heavy build;
• always when you ask for them.

The logic is cost and quality: a helper works with a fresh context, so it is cheaper and sharper on a self contained piece, and wasteful on a job that needs the whole conversation.

How a helper works

A helper gets one piece: a topic to research, a module to build. It cannot see your conversation, so Mori writes it a full brief with everything it needs. It works on its own, never stops to ask you, and returns a short result with file paths. Mori reads what comes back before she builds on it.

• Helpers can start helpers of their own, three levels deep.
• No more than three run at once on one machine. In Swarm mode, many small tasks fan out and wait their turn past that limit, and a helper that was turned away counts as a failure, never a success.
• No helper can change your memory.

Crews and the Captain

/crew and a task drafts a crew: a Captain and the members the job needs, such as a Builder, Reviewer, Tester, Researcher or Designer, each with its own brief and model. You see the draft and can change any brief, model or order; nothing runs until you press Go.

Members can each work in their own git worktree, one after another, and their work is merged back once at the end. Nothing is merged while your checkout has uncommitted changes. The answer is a crew report with each member's result and the Captain's check.

/captain runs a read only planning pass first and hands you a spec to approve before any building starts.

Lanes beside the turn

Send a separate request while Mori is working and it can start a lane: its own run in the same thread, with its own model and its own Stop, that answers in the thread when it ends. In a repository each lane works in its own worktree; elsewhere each file has one owner, so two lanes never write the same file. A finished lane's branch is merged only when you ask.

Agents you define

An agent is a short file: a name, a description of when to use it, its instructions, and optionally its model and tools. Mori ships with a Researcher, Coder, Critic, Designer, Explorer, Tester and Scout, and you can write your own (a security auditor, a copy editor) or install them from Community.

The logic of choosing one: there is no hidden scoring. Every agent's name and description is offered to the model as a choice, and the model picks one as an ordinary argument, with the whole request in view. That makes the description the routing rule, so it must say when the agent fits.

An agent whose definition asks for memory keeps its own notes between runs, one fact per line. Every note passes the same scan as any memory write, so a page cannot plant instructions in it, and a full notes file refuses new notes and says how to make room.

The Agents and Tasks page

Agents & Tasks shows what is running now and every job with several steps, live, with each specialist it handed work to. Agents from the community hub are listed there too. Long work that spans days is planned as a long task with milestones, and standing objectives live on the Board as cards that workers pick up.

Peers and more than one Mori

When you run several bots, one can hand a question to another that fits better. The peer answers inside both its own limits and the asker's, so a hop can never do more than either bot may.

On one computer you can also run more than one Mori: Run more than one Mori in Settings makes a copy with its own folder, sign in, settings and memory, for keeping work and personal apart. A copy starts without your keys; one button offers it the main Mori's keys when you want them there.

Agents outside Mori

Mori speaks A2A, the agent to agent protocol. Give her an agent's address and she reads its card and remembers it; then she can message it, carry on a conversation with it, or ask several agents with the same skill in parallel and merge the answers. Add one in Settings > Connections, under More ways to connect.

• Only public web addresses are allowed. Local and private addresses are refused, so a web page cannot point her at a service on your network.
• Another agent's reply is data, never instructions to Mori.

Mori can also be reached as an MCP server by other agents; see MCP.

Cost and control

Each helper, crew member and lane adds its own model calls to the turn, on your key or your plan's credits. The box in the turn shows each agent's model, tokens and time while it runs, and each has its own Stop. Helpers, crews and lanes run at the Standard and Full Agent permission levels. For the tools behind all of this, see Agents, helpers and peers.