The first line is the headline

Every step on foldrun ends with a reply from the model. That reply is stored whole as the step’s conclusion, and the first line of it is the headline: the line in the notification, the line on the run’s row in the dashboard, the line a person reads before deciding whether to open the run.

So the first line matters more than the rest of the reply combined, and a model left to itself will write “I have completed the analysis of the target keywords.” as the first line, every time.

The rule

We added a rule to every reporting agent, then to the runtime prompt itself: the first line is a verdict, with numbers, and no preamble. Not “Analysis complete” but 12 of 75 targets absent · 3 newly gone · 1 down 5+. Not “Here are my findings” but READY — 6 links, all on approved addresses.

Then we enforced it. verify: matches: ^(READY|NOT READY) on the reviewer step; a regex on the reporter that demands a count in the first line. A step whose headline does not match does not hand off.

The bug that taught us what a conclusion is

The regex failed on a headline that was obviously correct. The step’s reply began READY — and ^READY did not match it.

The reason: verify: was testing the step’s result, which at the time meant every turn of the conversation joined together — the tool calls, the intermediate thinking, and finally the reply. ^ anchored to the start of the first tool call. A (^|\n) workaround got the desk running while we fixed the platform.

The fix drew a line we should have drawn on day one. A step has a result — everything that happened — and a conclusion — the final reply. verify: tests the conclusion. each: lines of iterates the conclusion. The next step receives the conclusion. The headline is its first line. The result is on the run page for whoever wants to read the whole thing.

What earlier groups see

The same week we found the second half of the problem. A reporter in group three was writing confident summaries of measurements it had never seen, because the runtime handed each step the conclusion of the group immediately before it and nothing earlier. Group one’s seventy-five-row table was invisible by group three, and a fast model asked to “summarise the measurements” summarised what it imagined.

Now a step receives the conclusions of every earlier group, newest last, capped at thirty thousand characters. The reporter reads the table. The headline has real numbers in it.

Why we make this much of one line

Because supervision is done at the speed of headlines. A person with seven desks and eleven flows does not read runs; they read a column of first lines and open the two that look wrong. A platform that makes the first line reliable makes a whole team’s supervision cheap. One that does not is asking people to read everything, which means they read nothing.