everything the reading notices, turn by turn.
the topic, the task someone came to finish, the intents inside it, and the signal on each turn.
one conversation, read completely. then the same reading on every other one.
one conversation · product questions · enterprise, EMEA
nobody typed any of the labels on the left and right. they were attached while the conversation was still running.
one conversation shows you something.
the same reading, on all of them, settles an argument.
one reading, at three distances. the turn, the pattern that repeats across thousands of turns, and the person on the other end of both.
nobody phrases it the same way twice. search finds them anyway.
ask for “summaries written from an empty source” and every matching task comes back. the ones that said it plainly, the ones that only grumbled, and the ones where the tool event is the only trace.
- matched on intent, not on keywords
- plain english in, conversations out
- every result opens straight into its replay
528 conversations · ranked by meaning
enterprise, EMEA · task failed
“this is missing half of it.”
self-serve · repeated request
“do that again, with all of them this time.”
no complaint in the text · matched on the tool event
fetch_documents returned 0 rows, six times, then they stopped.
a conversation rarely goes wrong all at once.
it goes wrong on one turn. every turn carries its own score, so you get the sentence where it turned and the reply that turned it.
“this is missing half of it.”
turn 9 · two turns after 6 of the 9 documents came back empty.
- scored on the person's turns, so the agent's tone never flatters the number
- the turn where it changed is marked, so there is something to open
- one sour conversation is noise. the same drop on the same intent, across hundreds, is a fix
a task does not fail. one step inside it does.
task progress ties every request to the agent action that followed it. the task below has three steps. two of them finished, and the summary still came back half empty.
the same person, every conversation, in one line.
a person's page joins every session they have had. what they attempted, what finished, and which signal keeps turning up. open any point to read the conversation behind it.
session 1
summarise a thread · done
session 3
find the source · done
session 5
shipped summary · partial
session 7
shipped summary · failed
session 9
same task · failed again
three attempts at one task, and the same empty documents each time. read as three conversations, that is three shrugs. read as one history, it is one thing to fix.
cohorts are built from fields you already send.
plan, region, how often they come back, and any custom field the agent already attaches. combine a few and the group exists.
- saved once, then available to search, charts and monitors
- open a cohort to reach its people and their conversations
- behaviour works the same way: returning users is a field like any other
your agent is not failing everywhere. it is failing here.
conversations group into topics on arrival. each topic breaks into the tasks people attempt inside it. rank the topics by how often those tasks fail, and the argument about what to fix next is mostly over.
questions
and editing
a document
settings
task failure rate by topic · 12,400 conversations read, up 18% on the previous period · illustrative data
pick one signal. see which groups are carrying it.
a product-wide rate is an average of groups having very different experiences. plotted against that average, the same signal becomes a short list of who to fix it for first.
signal · documents came back empty · product-wide rate 11%
points above or below the product-wide rate · illustrative data
every claim in the summary links to the turn under it.
the summary names the task, the steps that finished, the one that did not, and the signal that decided it. read it out, then open the turn behind whichever sentence somebody doubts.
- written for a conversation, a person, a topic or a cohort
- the status comes from the turns, not from a thumbs-down
- nothing is claimed that has no turn under it
the conversation · 9 turns and 4 tool calls
- t1summarise what my team shipped.
- t2which space should i read from?
- t3the engineering space.
- t4found 9 documents in scope.
- t5the engineering space, not the whole workspace.
- t6re-read. 9 documents in scope.
- t76 of the 9 documents came back empty. summarising the other 3.
- t8summary written from 3 documents.
- t9this is missing half of it.
the summary
someone asked the agent to summarise what their team shipped. the source was corrected once, then settled. reading it did not finish: 6 of the 9 documents came back empty, and the agent wrote the summary from the other 3 anyway.
- task statusfailed
- failing intentread it
- decisive signaldocuments came back empty, turn 7
864 attempts stopped. this is where.
2,400 people asked for this summary and 1,536 finished. the other 864 stopped on one of three steps. the biggest one is first, and that is the order to fix them in.
- read it528 stopped · 22.9%the documents come back empty and the agent moves on. halve it and 264 more attempts finish. 64% becomes 75%.
- write the summary244 stopped · 13.7%the summary gets written. it covers what was read, not what was asked for.
- find the source92 stopped · 3.8%the agent picks the wrong space, or asks which one and never gets an answer.
read one conversation this way. then read all of them.
every conversation read on arrival, and a way to move between the pattern and the turn.
no email, no call. the playground is the live workspace.