Learn · multi-agent terminal
What is a multi-agent terminal?
By Lalla Asmaa Alaoui, founder of Mel · Updated
A multi-agent terminal is a terminal where several AI agents share one task, each taking a part such as planning, building or testing. They work inside your real project, under the approvals you set.
The word gets used for two different things. Some tools call themselves multi-agent because they let you open many agent sessions side by side, each doing its own job. Others mean a coordinated team: agents that share one goal and hand work to each other. Both are useful. The second is what changes how big a job you can hand over.
Why several agents instead of one?
Big tasks fail in the hand-offs. One agent holding a whole feature in one long conversation starts to lose early details, and its last change can contradict its first. Splitting the work fixes that the same way it does for people:
- Smaller briefs. Each agent gets one clear part, so its output is short and easy to check.
- Separate roles. The agent that plans is not the only one judging the result.
- Parallel progress. Independent parts (an API change, its tests, its docs) do not have to wait in line.
- Easier review. Several small, purposeful changes beat one giant diff.
The cost is more steps, and so more model usage. A team is worth it for many-step work, not for a quick question.
How a team of AI agents works
A team of AI agents is a group of agents that share one goal and split the work between them.
- Plan. One agent reads the task and breaks it into parts with a clear finish line.
- Build. Other agents take the parts: one changes the API, one writes tests, one updates the docs.
- Check. The work is run and tested, and the parts are checked to fit together.
- Report. You get the result and the changes to review.
Use a team for features that touch several files, refactors, migrations, bugs that cross modules, and research with several sources to compare. Use one agent or one chat for a question, a snippet or a one-line fix.
What agent orchestration means
Agent orchestration is the coordination behind a team: deciding which agent does which part, in what order, with what information, and how the results are combined and checked. It has two meanings. Orchestration as a framework is code libraries developers use to build their own multi-agent products. Orchestration inside a tool is a coding tool that runs several agents for you, so you see the plan, the progress and the result without writing any orchestration code. Mel is the second kind.
How Mel does it
- Orchestration is built in. Describe the task once; the team splits it, then plans, builds, tests and ships.
- No framework to learn. You see the plan and the result, with no orchestration code.
- More agents, same control. Every agent's commands and edits run under the approvals you configure.
- Beyond code, the same agents take on research, PDFs, photos and files.
How the approaches compare
| One agent | Many sessions side by side | Coordinated team (Mel) | |
|---|---|---|---|
| Who splits the task | The agent, step by step | You | The team plans it |
| Agents share one goal | n/a | No, each is separate | Yes |
| Checks the combined result | Itself | You | The team, then you |
| Orchestration code to write | None | None | None |
| Best for | One focused change | Several unrelated jobs at once | One many-step job |
| Your job | Approve, review | Juggle and merge | Set approvals, review the result |
How to keep control of a team of agents
- Set approvals before you start. Decide which commands and edits need your yes.
- Define "done". "Tests pass and the page loads" beats "build the page".
- Review the diff, not the summary. The summary is an opinion; the diff is what changed.
- Keep tasks reviewable. If a result takes more than a few minutes to read, the task was too big.
FAQ
What is a multi-agent terminal? A terminal where several AI agents work on one task together, each taking a part such as planning, building or testing.
What is a team of AI agents? A group of agents that share one goal and split the work, with the results combined and checked before they reach you.
What is agent orchestration? The coordination of several agents toward one goal: who does which part, in what order, and how the result is checked. In Mel it is built in, so you write no orchestration code.
Does a team of agents cost more? It can, because more steps run. Save teams for jobs that need them.
Can I stop agents from running risky commands? Yes. Mel's configurable approvals let you decide which commands and edits need your approval.
Mel – the agentic terminal · openmel.dev. A multi-agent, multi-model AI terminal for Mac, Windows, Linux and VS Code. Free with your own API key.
Lalla is a hands-on tech founder, agentic infrastructure architect and the vision behind Mel.
Follow on LinkedIn