All chapters
0:24

Building an agent#

You set one up once and then you just talk to it. That is the difference from a workflow: a workflow you lay out, an agent you hire.

An agent's workbench: the settings column, the conversation beside it, and the approval it asks for before acting outside.
  1. What an agent is

    «Agents» in the sidebar is this workspace's wall of them. Each one has its own model, its own tools, its own knowledge and a memory that carries between conversations. You build it once and use it many times.

    Do not confuse «Agents» with «Conversations». A conversation is an agent your customers talk to; an agent is one you talk to yourself.

  2. A new agent

    «New agent» makes one and takes you into it. Rename it at the top of the page and pick the model it thinks with. Until you do, its card on the wall says so.

  3. The master prompt

    The most important field in the settings column. In your own words, write what this agent is for, what matters to it and what does not. That text is in front of the model in every conversation, from the first sentence.

    Short and specific beats long and polite. «Business news editor; cuts any piece to five lines and never rounds a number» does more work than a page of general description.

  4. Tools, and what it may do with them

    «Tools» is what the agent may do outside the conversation; «Permissions» is whether each of those happens on its own, asks first, or never. Any single tool can have its own exception.

    A new agent starts on «reads are free, the rest asks»: anything that only reads happens by itself, anything that writes or sends asks you first. Leave it there until you have watched it work.

  5. What it knows

    Three separate things: «Knowledge» is documents from the workspace library it can search, «Files» is the agent's own folder, available in every conversation, and «Memory» is what it has remembered about you and your work.

    You can write and delete memory yourself. It is one file belonging to the agent rather than one per person: everybody who talks to it reads the same one. The conversations themselves work the same way.

  6. Working when you are not there

    «Workflows» hands the agent boards from this workspace so it can run them as tools, and «Schedule» is where you tell it to work on the clock and where to deliver what it made.

    Under every answer it says what that turn cost. The credit comes out of the workspace, not out of you.

  7. You do not have to fill it in by hand

    Every one of those fields can be asked for in the conversation itself: «From now on send the weekly report on Mondays.» The agent proposes the change, shows «now» beside «after», and nothing moves until you press «Apply».

  8. Recording a version

    What you see in the settings column is the draft, and that is what the agent is talking to you with. «Record a version» keeps a photograph of it, with a note you write, so you can come back to it later.

    Restoring a version restores the settings only. Files and conversations are untouched, and stored secrets stay the ones you have now.

Build one and put it to work in your first sentence.

Create an account

Revaal

Manifesto

A person’s attention is the most valuable thing they bring to work. Every one of us has a limited amount of energy, focus, and presence each day, and a day that is spent does not come back. This is true for every single person on a team, not only the ones at the top. So the real question for any company is a simple one: what are we asking people to spend their days on?

Too often, the answer is repetition. We ask people to copy, paste, reformat, check the same list again, and write the same message for the hundredth time. The work gets finished, but something is lost. A person doing repetitive work is a person who is not thinking. We believe this is a serious mistake, not because such work is beneath anyone, but because a human being is the wrong tool for it. Machines are patient with repetition. People are not, and they should not have to be.

We are building for the other side of that trade. Models now have the processing power and the intelligence to carry the repeating part: the reading, the sorting, the drafting, everything that follows a known pattern. Hand that over, and look at what is left for people: creating things, staying curious, wrestling with problems that have no certain answer, sitting with a question while it is still unclear, and making something new and then changing it again. That is what a human day should be spent on. We think it should be the normal day, not the rare one.