Skip to content

Several agents on the code, each with its model.

The code team is a coordinator and agents. You choose each one’s model, the best for its job: one that reasons at length, a fast one, one that can look at screens. The coordinator hands the work to whoever is needed.

Book a demoSee the presentation

A model for every job.

Each agent has a name, when to use it and its model, taken from the provider’s list of models, which is re-read every hour. The hard task goes to a model that reasons at length, the touch-up to a fast one, and when a screen needs looking at, to one that can.

  • At most twenty-four agents.
  • An empty model is the coordinator’s.
  • Only models that use tools.

The coordinator hands the work over.

Each agent becomes a tool of the coordinator, and its “when to use it” is the description the coordinator reads. The agent works in a session of its own, with its model, without seeing the conversation. The result comes back, and at the end the coordinator says who did what.

In detail

  • Who coordinates.

    You choose its model. Left empty, it is the default of whoever installed Fallthere.

  • Five starting points.

    Code, frontend, reasoning, images, fast: ready, without a model. You put yours in, and change them as you want.

  • Each one’s instructions.

    For each agent the name (two to sixty characters), when to use it, the model and the instructions, up to four thousand characters.

  • It can hand work over.

    With Can hand work to others in the team the agent receives the team. Without it, it works alone.

  • It must look at screens.

    With Must look at images and screenshots the model must be able to: one that cannot is not saved.

  • A rule, not a guarantee.

    In front of the task there are five rules: what fits in one line “you must hand over”, independent parts go together, and at the end it says who did what. A model can ignore them.

  • Some limits.

    At most two handoffs in a row, ten minutes per part, a task of up to twenty thousand characters. Stop stops everyone.

  • The spend is yours.

    It uses OpenRouter models with your Pan key, and your cap, like Pan in the column. Never the organization’s key.

How to use it

Five moves, from the team to the task.

  1. Turn on Coding.

    And connect GitHub, if you have not. How.

  2. Open The code team.

    In Settings, Account, The code team (it is there only with Coding on), or the The team button in Pan on code.

  3. Choose who coordinates.

    Who coordinates, The coordinator’s model. Or The default one.

  4. Add the agents.

    Add an agent: the name, When to use it, the model. Or Or start from one of these. Then Save.

  5. Give it a task.

    In Pan on code, What should I do in the code? and Ask. The coordinator hands the work to whoever is needed.

Examples

Made-up people and data.

  1. Luca, Mirto Labs

    Frontend to a model that looks.

    Luca creates the “Frontend” agent with a model that can look at screens, and ticks Must look at images and screenshots. The coordinator hands it the interface changes.

  2. Luca, Mirto Labs

    The touch-up to a fast one.

    For small fixes there is the “Quick” agent, with a light model. It answers at once, and the spend on Luca’s cap is small.

  3. Marta, Mirto Labs

    Who did what.

    At the end of the task the coordinator writes who did what. In Changes Marta sees the files changed on the agent’s branch.

Connects to

What you do here shows up elsewhere, without redoing it.

Coming soon

No dates: it arrives when it’s ready.

  • Group coding: several people in the same conversation with the agent, one turn at a time

Frequently asked questions

Can I use a different model for each agent?

Yes: that is the point. Each agent can have its own, and an empty one uses the coordinator’s.

Which models can I choose?

OpenRouter’s that use tools, from the list re-read every hour: new ones show up on their own. AI models.

Do agents see the conversation?

No: they work in a session of their own and receive the task. The result goes back to the coordinator.

Who pays?

Whoever uses it, with their OpenRouter key and their cap. Coding’s consumption is estimated low: four characters per token.

Does the coordinator always hand work over?

It is pushed to with the rules in front of the task, but a model can ignore them. That is why the team works better with a coordinator that follows instructions.

Is the team mine or the organization’s?

Yours: agents are per person, and everyone makes their own.

Let’s see it together.

We show it to you live.

Book a demo

Your mail app opens with a message already written. Or write to info@firmamentotechnologies.com.