Skip to content

Projects, agents and tasks

Projects — a shared place for one topic

A project keeps chats, files and finished results together in one place. Give a colleague access to the project and they get all of it at once, rather than being sent files one by one.

Start a project as soon as a topic outlives a single conversation, or as soon as somebody else is involved.

  • Create one in Projects, using New. Write a description — it tells your colleagues what the project is for.
  • Upload the files to the project, not into one chat. Then every chat in the project can quote them and every member can open them.
  • Give a colleague access from the project page. You can only share inside your own organisation. They then see the project's chats, files and results.
  • Start the chat inside the project. It can use the project's documents without you switching anything on, and /memory save in it stores the fact for the whole project (see Memory and commands).
  • Connectors you activate on the project apply to the project's chats by default — see Connectors.

Results the assistant produces for you — a document, a spreadsheet, a diagram — are kept with the project too, so they are not buried in a conversation.

Deleting a project deletes its chats, its files and those results.

Agents — a helper you set up once

An agent is a set of standing instructions with a name. Instead of explaining the same job every time ("check this submission against our review criteria and tell me what is missing"), you write the instructions once and then call the agent by switching it on.

Some agents are built by your organisation and shared with everyone. You can also make your own.

Using one

Open the Tools sheet in a chat and find the Agents section. Each agent has three positions:

Position What it means
Auto The agent is available, and is used when the question calls for it. This is the starting position.
Always Keep this agent at the front of the queue for this chat. Use it when the agent is the point of the chat.
Off Never use this agent in this chat.

Your choice is remembered for that chat. You do not have to address the agent directly; the assistant brings it in when it is relevant and folds its answer into the reply.

A worked example

Your organisation may already have the Architecture Board pair installed.

AB Project Validator checks whether a project has been set up properly: does it have an identifier, a description, stated business benefits, a named owner and a defined scope.

Architecture Board then judges a submission. It accepts only when the project foundation is sound and the submission explains its effect on the systems already in place, describes the technology it will use, and states the costs. The verdict — accepted or rejected, with a reason against each criterion — is written into the chat, so the review is a conversation instead of a form.

Making your own

Go to Agents and choose New. You give it:

  • a name your colleagues will recognise;
  • instructions: who it is, what it should check or produce, and how the answer should be laid out;
  • optionally a specific model, if one suits the job better than the default;
  • optionally the connectors it may use.

Share it with the whole organisation when colleagues should have it too.

Tasks — work that takes several steps

Some requests are really several jobs: "write the story, draw the diagram, then build the slides". For those, the assistant sets out a plan — the steps in order, with the dependencies between them. A step that needs another step's output waits for it. Steps that do not depend on each other run at the same time, which is faster than doing them one after another.

An ordinary request is not turned into a plan. One question, or one file conversion, is done directly.

Watching a plan run

The plan appears in the chat as a card. Its heading counts where the work stands — 3 running · 2 queued · 1 done — and one line beneath that says what is happening at this moment, so a glance tells you whether to wait or come back later.

Open the card and you see every step. Each one carries its own activity line while it works, the same kind described in Chat basics, and reports what it produced when it is done — wrote 412 words for a step that wrote something, or a file you can download. A step that runs into a momentary problem with the model supplier shows its retries instead of going quiet, and keeps trying for up to about ten minutes before giving up.

The Tasks panel is the same information away from the chat. You can watch progress there, jump back into the chat that started a plan, and stop work you no longer want. Clear finished clears away completed lists, and a single list can be put away with the archive icon.

When a step fails

The card turns a warning colour and opens itself, so a failure is never left folded out of sight. Its heading says what happened overall — Plan failed — 3 steps failed, 2 of 6 done — and inside it:

  • the step that went wrong is marked Failed, with a short reason;
  • every step that was waiting for that step's output is marked Blocked, with a note naming what it was waiting for: "Skipped: this step needs the output of …, which did not complete."

Everything that did succeed is still yours. Finished steps keep their results, and the files they produced stay with the project. Text that had already appeared in the conversation is real and stays there — a step failing later does not take it back.

What the assistant will no longer do is finish a document from an incomplete set of ingredients. If you asked for four drawings and a slide deck, and only one drawing was produced, the deck is marked Blocked rather than being built from the single drawing and handed to you as though it were finished.

That is a deliberate trade, and it is worth stating plainly: you will sometimes get fewer finished files than you would have before — but you will not get a file that quietly leaves things out. A deck missing three of its four drawings still looks finished, and the gap tends to be found by somebody in the room rather than by you. A step marked Blocked tells you at once what is missing and which part to ask for again.

To carry on, ask for the missing piece in the same chat. The results that did succeed are still there to build on, so you are not starting the whole request over.

Work that runs on a schedule

In Tasks, open Scheduled tasks and choose New to have an agent run without you being there — a Monday morning digest, a weekly check.

You give the run a name, pick the agents it should run, and choose how it starts:

  • Schedule — it runs repeatedly. The timing is written as a cron line, the standard five-field pattern used by scheduling tools: minute, hour, day of month, month, day of week. 0 9 * * * means every day at 09:00. 0 8 * * 1 means Mondays at 08:00. You also pick the time zone it should follow, so a daily run happens at 09:00 where you are.
  • Manual — it never starts by itself; you run it when you want it.

Two further options, marked MCP and A2A, let another program or another agent start the run. Both need technical setup by whoever built the integration; if you have not been told about one, you do not need it.

A scheduled run happens when nobody is watching, so it deliberately does not use connectors that depend on a person signing in with their own account. Those are skipped rather than borrowing anyone's access.