The most common complaint about AI tools is that they do not say what they did. A text arrives, it looks fine, and you have no idea what it rests on. If you do not like it, you do not know where to fix it either.
In Orbi an agent's work shows in four separate places inside the card. Knowing which one to look at takes the guesswork out of working with an agent.
The checklist says how much of the job is done
The agent splits a long job into steps itself and writes each step into the card's checklist. You do not add the steps. The agent opens them and ticks them off.
Each step keeps its status and a short summary of its output. You look at one step and understand what was done without reading the whole thing.
Steps are not deleted. A failed step is not removed from the list, it is marked as failed. That way "tried and did not work" does not get confused with "never tried". The difference between the two is very useful when handing out a job a second time.
You can see where it stands without opening the card. The checklist's progress sits on the card as a segmented bar and a counter, so glancing at the board shows how far each card has got without opening the list.
Do not mix up output and comments
A card has two separate writing areas and they do different jobs.
The Outputs area is for delivery. The final content the agent produced goes there. A report, a table, a text, whatever was produced.
Comments are for communication. Which assumption the agent made, what it could not find, the question it is asking you, all land there.
Knowing this split lets you read a card quickly. If you want the result of the work, you look at the output. If you want to know how it went, you read the comments.
History, the whole life of the card
The History tab shows a single timeline, and two kinds of record flow through it together.
Activity entries say what happened to the card. State changed, assignment changed, a comment was written or deleted, a step was added to the checklist, an output arrived. Each line says who did it and when.
An agent run is a single line. "Worked on this card, updated three records, moved the card to review" tells the whole run. If every write inside the run were its own line, a ten-step run would fill the list on its own. That detail sits under the line and opens when you ask for it.
Run lines are recognised by the thin purple line on their left. The line means "an agent did this". If the run failed the line turns red, because what needs saying there is no longer who did it but how it ended.
The diagnostic trace answers the technical question
When an output comes back weaker than expected, the questions are known. Which agent ran, with which model, was there a fallback to a backup model, was a sub-agent called, how many tokens were spent.
All of that sits under the agent round. So you answer "was the model weak or was the brief thin" from a record rather than a guess.
The same trace helps on the cost side. If a card cost more than expected, you can see which round spent how much.
When an agent touches a connector, the service's own mark sits next to that step, so you can see which system was contacted without reading the line.
The same list does not live only on the card
The list on the card's History tab is drawn in the same language in four more places. On an agent's own page only that agent's work, on a collection screen what happened in that collection, inside a record that single record's history, and on the Activity tab of the Inbox the whole workspace.
The last one is the first place to look in the morning. "What did the agents do overnight" gets answered in one list, and one of the views filters straight down to what needs attention, meaning failed runs and cards that landed in review.
History cannot be changed
The activity log is not deleted and not edited. Even when a comment is deleted, the log keeps a "this comment was deleted" entry along with a text summary.
For the sentence "something was deleted" to mean anything, what was deleted has to be known. A log that also deletes the trace of what was deleted stops being an audit log.
Automatic moves that cannot be attributed to anyone are not lost either. They show as done by the system. A state change triggered by a flow, or a transition made because a ceiling was hit, is marked that way.
Retention runs on two clocks. The step trace inside a run is kept for 180 days, the run itself and standalone actions for 400 days. So "who did what" can still be answered months later, while the "what happened at which step" detail is cleared earlier.
Asking for a revision does not start over
When a card lands in review, you say what you do not like in a comment and put the card back in the queue.
The agent does not redo the job from scratch. Steps that sit completed in the checklist are not repeated. Only the relevant step runs again. Saying "the third section is weak" does not cause the whole piece to be rewritten.
If two people, or one person and one agent, edit the card at the same time, the conflict is not silently overwritten. It gets detected.
If it is code work, the PR shows on the card too
If a pull request references the card, a mark sits in the card's top strip and its colour says the PR's state. Grey means open, green means merged, red means closed without a merge. The same mark appears in the list view, so you see it without opening the card.
A merge does not close the card. Marking a card done is a human decision. When the PR is merged, only the mark turns green.
Where to start
Before you approve the next agent card, look at three places in order. First the checklist, to see how many steps the job took and whether any is marked as failed. Then the output, to read the delivery itself. Last the comments, to understand which assumption the agent worked from.
All three take a few minutes and take the "can I trust this" question out of the realm of guessing.
The checklist, history and the revision loop are covered in detail in Tasks and Kanban in the help center. To see the whole workspace in one list, see Inbox, and for how work gets split, Orchestrator.
If you want to try it, you can open a free account and hand your first card to an agent.