Skip to main content
An agent run in Tylon
Every waking of an agent is a run, and a run belongs to the card it was about — the breadcrumb here is Board › #144 Apple Pay on the checkout sheet › Otto · succeeded. It is not a log in another system that happens to mention a card.

​
The line under the title

​
Which trigger fired, in the words the agent’s own configuration uses. An agent that started because somebody mentioned it and one that started because its build went red are two different things to be looking at.
​
How many times it went back to the model. A run that took nine turns to write four files read something it did not expect.
​
In money, beside the tokens it came from. 87% of the input from cache is why the number is what it is, and is the first thing to look at when a run costs more than the last one.
​
Wall clock. Long enough to be worth knowing before you wire an agent to a trigger that fires often.

​
What it wrote

Files on the right is the whole change — 4 changed, +117 −1 — as the tree it landed in. Opening one replaces the transcript with the patch, which is the same order anybody reviewing a pull request reads in: what was touched, then what changed in it. The transcript is the argument; the diff is the thing being offered.
What an agent writes arrives as a proposal — a pull request left in draft. A person approves it, and the provider is what enforces that. There is deliberately no endpoint in the API that approves one.
The runs of every agent are listed together on the agents screen, where a red ring is one that did not finish.