Changelog

Your agent can build its own colleagues

Until yesterday, if your agent said «this one is a job for an image-maker», the rest was yours: leave the conversation, go to the roster, make one, pick its model, write its instruction, come back, and add it to the list of colleagues by hand. Now the sentence is enough — «make me an image colleague that draws the covers for the news» — and it is created, its instruction written, and taken on as a colleague there and then.

The colleagues you already have change the same way: «shorten the editor's instruction», «put it on a different model», «give it this skill». You do not need to know which column is where.

An agent's memory is a file now: memory.md

Until yesterday, everything an agent learned was a «note», and before writing one down you had to pick its kind (preference, fact, decision, context) and its scope (about me, about the work) from two dropdowns. Nobody wants to classify a sentence. Now it is one file — memory.md — that the agent writes into from what you say, and that you open and edit like any other text.

The file belongs to the agent, not to one person. What one colleague teaches it today, the next colleague has tomorrow: the channel, the tone, the decision taken last week. Saying it once is enough, for the whole team.

The agent keeps working after you close the page

A turn belongs to the conversation, not to the tab it was started in. Close the page, come back later, and the answer is waiting. Work that takes minutes can be started and left, and the agent no longer stops just because you did.

The wall of agents shows which one is answering somebody right now, so finding out does not mean opening each of them. And a file the agent writes opens beside the conversation that wrote it: no download, no round trip.

• What a turn cost is written under every answer, in Toman and at the rate it was paid at. A team's spend reads where it happened rather than at the end of the month.
• The order of conversations is dragged into place and kept on your account, so the same order is there on whatever machine you sign in from.
• Focus clears the page of everything but the conversation, for when you only want to read and write.

The agent gets a workbench of its own

The agent used to be a handful of pages under the workspace. It has a page of its own now, at /agents/<id>: everything about an agent is there, and the workspace keeps only the lists.

Conversations get a column of their own: rows that say a name and a time, and a menu on each for what can be done to it. Tools no longer buries the rest of the settings under itself either.

• Sharing An agent is shared the way a board is, with the same roles. An agent one person built works for their team.

Build an agent by talking to it, and let it remember

Building an agent is no longer filling in a form. You say what the job is, and the agent proposes its own configuration: tools, permissions, skills, subagents, documents and triggers, each as a card you approve where you are.

Memory comes from the same place. Notes the agent keeps between conversations, one set per person and one shared, folded into every turn. What you told it last week does not have to be said again.

• Versions Commit a frozen copy of an agent, restore one, and pin automation to the version you tested.
• Boards as tools An agent can call your own workflows, and it can build one.
• Other modalities An agent that makes pictures, clips and speech, reads speech, and lends those to the agents above it.
• Dictation The composer dictates over the pocket socket, so nothing has to be typed.

The agent: configure it once, come back to it

Until today every agent was a step on a board and ended with that board. An agent is one that gets saved: you configure it once, it has a name, and you come back to it. Its configuration stays with it rather than with the run it was used in.

The master prompt is no longer a field you scroll inside. It is a file you open.

• search_web The agent can search the web.
• read_url And it can open a link and read it.

A re-run picks up from the step that failed

A failed run used to mean starting over. A re-run now continues from the step that failed, and everything built up to it stays. An eighth step that errored no longer pays for the seven before it a second time.

Inputs take documents too: pdf, docx, markdown and the rest. Every input opens, and its rows can be read and edited, without fixing the file somewhere else and uploading it again.

• Billing says which workflow a cost belongs to, not only how much it was.
• How it got there shows the reader the path the agent took, so a strange output can be traced.
• The primary workspace cannot be deleted, managing one has a way out, and ownership transfers from the admin panel.

Plugins: a shelf of every ready-made tool

Every tool a step can be given now has a shelf, categorised and searchable, with a «new» mark on what has just arrived. A service with two ways in appears once, with a note naming both, rather than as two rows one of which is wrong.

The library became a file manager: make folders, upload, rename and move. The brand kit got a tab of its own as well.

• Connect asks which client you have first, then gives you that client's three steps. The consent screen shows both ends of the connection.
• The assistant sees the board's context, and its tools no longer go without credentials.