Long agent work — watch it, steer it, stop it
When an agent works on something long, the team watches it in one reply that fills in as it goes, can add to or correct the task while it works, can stop it, and always gets an honest answer — never just "it ran out of time". Mockups of every moment, what changes from today, and the order it ships in.
Why
Agents now run out of time often. When that happens today, the member gets "I couldn't answer: it ran out of time" — and the work the agent already did disappears from view, even when most of it was finished. Meanwhile, nothing the team says while the agent works reaches it: a second message starts a second agent run that knows nothing about the first.
People talk to an agent the way they talk to a colleague: they ask, then add "oh, dark mode too". This design makes the agent behave like that colleague.
What it looks like
Tap through the moments. The example is Marcie fixing the status colours in #design, with Turbo and Vee adding to the task while she works.
The reply appears as soon as the agent starts and fills in as it works. The line with the dot is what it is doing right now.
# design
Turbo 11:48
@Marcie fix the status colours, esp orange and green
Marcie AGENT working · 0:42
Checking all six status colours against white and dark backgrounds.
Orange is 2.9:1 and green 3.1:1 on white; both need 4.5:1. Fixing those two first.
Updating green…
What changes
| Moment | Before | With this design |
|---|---|---|
| The agent starts | A grey list of tool names appears | One reply appears and fills in, in plain sentences, with a line showing what it's doing now |
| Someone adds to the task | A second, separate run starts and knows nothing of the first | The same run takes it in at its next step, answers it in the reply, and the reply moves below it. Sent once the agent has begun its last step, it waits for the answer and starts the next run |
| Two people disagree | Two runs, two answers | The latest message wins, and the agent says so — anyone can correct it |
| The agent finishes | The tool list is replaced by the answer | The answer replaces the progress; How I worked opens the steps in plain words |
| Time or steps run out | Built: a wrap-up — what was done, what's left, and "reply continue" | The same wrap-up, as the living reply's ending |
| You want it to stop | Nothing to press | Stop above the composer, for the person who asked; the agent wraps up the same way |
| A message after it's done | Starts a new run | Unchanged: starts a new run with its own reply, which can read the last one |
The rules, in one place
- One conversation per channel. Whoever writes to a working agent reaches the same run — there are no private side-conversations with it.
- What reaches a working agent is what would call it today: a mention, a reply to one of its messages, or a channel rule (such as filing what's dropped). No new setting, and no "busy mode" to pick.
- Every kind of run works the same way — a question, a dropped file, a scheduled routine.
- While it works, the reply's position shows what the agent has read. Messages above it, it has read; messages below it, not yet. There are no "read by" marks. The posted answer takes its place in time order, so it can sit below chatter the agent never read.
- A message sent once the agent has begun its last step (its final thinking or its answer) waits for the answer, then starts the next run, which can read that answer.
- Anyone in the channel can add to the task. What the agent changes is still done for the person who asked, and only they approve an action. Only the person who asked can stop it, until workspaces have admins.
- Developers keep the raw steps. In developer mode, How I worked shows the tool names
(
edit_page,read_skill) instead of plain words.
What it won't do
- Make agents faster. The work takes as long as it takes; you just see it.
- Undo what an agent already changed. Stopping ends the run; writes it already made stay, and the wrap-up says which.
- Give each person their own conversation with the agent in a shared channel.
The order it ships in
| # | Step | What you can do after it | Where it is |
|---|---|---|---|
| 1 | Wrap up when time runs out | Get what was done and what's left instead of an error; agents get twice the time and steps | Built (PLAT-37), in testing |
| 2 | The living reply | Watch one reply fill in while the agent works | Built (AGENT-6) |
| 3 | Steer a working agent | Add to or correct a task while the agent works | Built (AGENT-18) |
| 4 | Stop | Stop an agent you asked | Designed |
| 5 | Keep going | A job too big for one run carries on into the next | Idea — waits until we see real demand |