TOOLS ARE VISIBLE NOW. Clicking any agent — in the trail or the "Worked
by" strip — opens a card saying what it is, when it runs, what it can
reach and what it reads. The answer is short and it is the reassuring
kind: four of the five hold no tools at all and reason from the policy
wordings, and the only tool in the roster reads a partner registry.
That answer stopped being true once, silently. Intake held a tool that
could telephone customers, its charter never mentioned it, and it used it
on every lead believing it was a duplicate check (89). Nobody could see
that from any screen. This is the screen that would have shown it — which
is the actual argument for building it, beyond a demo looking better.
Where an agent has no tools the card says so in words rather than showing
an empty section. "No tools" is the single most reassuring fact about
something that writes into an insurance file; a blank reads as missing
data.
The roster is a hand-maintained mirror of the employee config, like
STAGES and ACTIONS, and carries the same hazard: a tool added there and
not here is described wrongly, silently. Named in the file.
THE CONVERSATION WAS EATING THE TRAIL. Every WhatsApp turn was its own
entry — heading, actor, quote box — so a four-message exchange occupied
more of the trail than the entire underwriting chain, and a resend
printed the same sentence four times because the customer sent it four
times. A contiguous run now collapses into ONE entry: a WhatsApp mark, a
count, the latest line, and a link that opens the thread. A second
exchange later in the lead stays separate, because that is a different
episode in the story.
And the actor is right. The channel performs these as the system, so the
trail read "System · Customer Reply" — the opposite of what happened.
It says "the customer" now, in the channel's own green, and the rail dot
takes that colour so a scan shows where the conversation was without
reading a word.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Five of the six asks. The sixth — the whole row as the click target —
shipped in 704afe1 and is already live on both grids.
THE AUDIT TRAIL WAS FROZEN. Timeline fetched the rows once on mount and
never again, so a lead being worked by five agents in three minutes
showed the trail as it was when the page opened — on the one screen whose
job is watching work happen. The rows move to the lead page, which
already polls every 12s, and Timeline becomes presentational. One fetch,
one poll, and the conversation view reads the same rows so the two cannot
disagree.
THE ACCEPTANCE BUTTON OUTLIVED THE ACCEPTANCE. Recording an acceptance
twice is not a loop, a recovery or an alternative — it is meaningless,
and offering it invites someone to overwrite a customer's WhatsApp
acceptance with a worse record of the same event. Gone once
acceptance_ref is set, replaced by a line saying who accepted and when.
THE AGENTS HAD NO FACE. Five of them carry this workflow and the console
named them three different ways — "ai_engage" here, "Engage" there,
"Engage AI" elsewhere — so nobody could see that the thing which called
the customer and the thing which wrote the quote were one worker. One
roster now (api/agents.js): a short name, a colour and two initials each,
used wherever an agent is named. People get a disc too, in grey — a trail
where the machines are decorated and the humans are plain text reads as
though the machines are the important ones, which is backwards on a
screen built for oversight. A "Worked by" strip above the trail shows the
team at a glance.
THE CONVERSATION WAS A LOG, NOT A THREAD. Each turn sat as its own entry
among twenty others. Now one button opens it as a thread — theirs left,
ours right, oldest first — and the footer says plainly which parts it
holds: the quote, the read receipt and the acceptance confirmation are
sent by trigger nodes and never written to a field, so they are not
there. A thread that quietly omitted them would be worse than one that
admits what it is.
CONFIRMING A PREMIUM IS A SIGN-OFF, NOT A TASK. It is the only step
locked to one role and the only one that touches money — an employee that
could mark a premium received could put a customer on risk for a policy
nobody paid for. It sat fourth in a list of six queues. It gets its own
band now, shown only to the role that owns it, amber only when something
is actually waiting: a permanent alert colour on an empty queue teaches
people to stop seeing it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Engage can now answer a customer's question from the policy wordings
(82), so the lead file has a conversation in it rather than a single
recorded reply.
The audit trail renders both halves — "The customer said", then "We
replied" — so the exchange reads as one, which is what an operator
reviewing an acceptance nobody in the office made actually needs.
The lead file gets a Conversation group beside Proposal: the question,
the answer, and the count. Acceptance stays in Proposal, one tab away
from the words that produced it.
"Reply to the customer" is offered to ops_admin only. Engage declines
anything it cannot answer from the wordings, anything that is a
negotiation, and everything past three answers — so a person needs a way
to take over, and it has to be through the same thread or the reply
lands somewhere the lead file never sees.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Recording the acceptance was the third human touchpoint, and it was a
person in the office typing what a customer had said somewhere else.
The customer's WhatsApp reply now drives it (81), so this stage has no
step of its own.
"Record the acceptance" drops from a step to a folded recovery lever —
kept, because a customer who says yes on a CALL still needs somebody to
record it, and because a reply Engage judged ambiguous has to be
actionable by a person.
That left Quote Presented with an empty screen at exactly the stage an
operator is most likely to wonder whether something has broken, so the
waiting state now says so. Grey, and no pulse: nothing is happening, and
that is the correct state — unlike the AI strip, where a still dot would
mean something is wrong.
The reply itself surfaces in two places: the audit trail carries it as
"The customer said", and the lead file files it under Proposal beside the
acceptance it produced. The evidence and the decision it justifies are
one line apart, which matters when nobody in the office made it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
THE SIDEBAR NOW CARRIES COUNTS. It deliberately did not, on the argument
that a tally could only come from a second full list call that would then
disagree with the queue's own total. Right about the cost, wrong about
the conclusion: with no numbers the only way to learn whether anything
was waiting on you was to open all nine queues in turn, which is the
question navigation exists to answer.
There is no second call. The overview's existing fetch moved into a
PortfolioProvider that both surfaces read, so this is one call fewer than
before and the two agree by construction. Bounded at 200 rows, as the
overview always was; past that the honest answer is an aggregate
endpoint, not a bigger limit.
What the badge counts is what needs a PERSON — phaseOf again, so a lead
whose documents are in and whose AI chain is running is reported beside
the badge rather than inside it. An amber dot marks a queue holding a
renewal inside a week, which is the only reason to open one queue before
another and was previously invisible. Zero is shown rather than hidden:
"nothing here" is an answer, and a queue that disappears when it empties
makes the sidebar move under the cursor.
Closed is folded into a summary. Three of the twelve queues, opened about
once a week, and at equal weight they made the live ones harder to find.
REASONING OPENS IN A DIALOG, NOT INLINE. The audit rail is 320px wide
with one entry per step, and a 1,500-character rationale expanding in
place pushed a lead's whole history off screen to read one sentence of
it. Nobody reads a paragraph in a sidebar. The finding — the AI's own
first sentence, which is already the conclusion — stays on the line; the
working is one click away and one Escape back, headed by what is being
read ("Call — Log Contact") rather than by nothing. Escape closes, the
page behind does not scroll, and focus returns to the button that opened
it so a keyboard reader keeps their place.
Two headings that described the container rather than the contents:
Record → "Lead file", and the audit trail's subtitle now says what a
reader gets from it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Traced the console against the workflow it renders. Five gaps, all of
which made a working chain look like a broken one.
THE AUDIT TRAIL WAS READING THE WRONG KEYS. An audit row's data is keyed
by the ACTIVITY's field ids, which carry a per-form suffix — the call
notes arrive as contact_notes_2, the document request as
documents_notes_3 — and the timeline matched the unsuffixed global ids.
Almost nothing ever matched, so the trail was a list of activity names
with the agents' reasoning invisible behind it. The one narrative line
that did appear was an accident. It now reads the platform's own typed
`fields[]` (label, data_type, value) and matches on the base id.
That is also what lets it answer the question that was asked: an upload
now renders as "3 documents received" with each one named and openable,
instead of a bare "Collect Documents".
DATA_UPDATE ROWS WERE CLASSED AS AGENT WORK. A bookkeeping row inherits
the roles of whoever caused it, and the AI check ran first — so an AI's
field write appeared in the trail as an entry titled "Data updated",
while the toggle underneath still offered to reveal the others. The
activity id says a row is bookkeeping; the roles say who triggered it.
DOCUMENT PENDING IS TWO SITUATIONS. Before the upload a person has to
act; after it the lead stays in the same state while three AI steps run.
Both rendered as "waiting on the partner agent", so a lead that had just
been served showed an Upload documents button under a panel saying the
documents had been received, and counted against the Action-required
queue. phaseOf() derives the difference once, from documents_status, and
the header, the queue count, the row status and the action list all read
it. The upload demotes to recovery — "Replace or add a document", folded
away with the other levers.
A STOPPED CHAIN LOOKED IDENTICAL TO A RUNNING ONE. The rating engine
refuses to price without an IDV and writes so into quoted_breakup;
Engage then declines to raise a quote it would have to fabricate. Both
are right, and nobody was told: the refusal sat in a field on a tab and
the lead never moved again. blockedOn() surfaces it as an amber strip
that names the missing value and opens the form that carries it.
THE STALL MEASURE WAS DATED FROM THE WRONG COLUMN. progress() fell back
to created_at when updated_at was absent — which it always was, because
neither view returned it — so every lead older than half an hour would
have reported stalled. It reads updated_at only; the view supplies it as
of 76.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two things on the lead page.
THE AUDIT TRAIL REPEATED ITSELF. "Qualify Lead" appeared three times in
a row and "Partner Agent Lead" twice, which reads as the AI having done
the same thing three times. It had not. One submission writes several
rows, told apart by execution_state:
zk-act-qualify TRIGGER_PERFORMED the trigger's commit
zk-act-qualify TRIGGER_PERFORMED again, on the settle path
zk-act-qualify zk-state-qualified the one that moved the lead
Only the last means anything to somebody reading the file. They are now
collapsed into one entry carrying the earliest timestamp — when the
operator acted — and whichever state and payload is populated.
Matched on (same activity, within fifteen seconds) rather than on
execution_state, so a genuine repeat survives: Collect Documents really
is performed twice on a lead whose first upload was short, and those are
minutes apart. And matched by looking BACK through recent entries rather
than at the previous one, because the platform interleaves a DATA_UPDATE
row between the two halves of a submission — the rows to merge are near
each other in time but not adjacent in the list.
THE AI PARAGRAPHS HAD NO SUMMARY. attribution_reason and
eligibility_reason run to a paragraph each and opened folded, so the
finding could not be read without expanding. ClampText now lifts the
first sentence out as a headline and puts the rest behind "Show the
reasoning" — no new field, and nothing invented: the employees already
write a conclusion and then its evidence.
It falls back to plain folding when the split would be useless — no
sentence terminator, a first sentence that is the whole paragraph, or
one short enough to be a fragment rather than a finding.
The other half of that is in the charter (70_verdict_first.sql). The
headline is only as good as the sentence, and Intake was opening with
"Tool 29601 confirmed..." — an internal id — because it had been told to
"name the evidence, not the conclusion". Right about content, wrong about
order: it now leads with what it decided and gives the evidence second.
Three problems, and the first one is the reason the other two were hard
to see.
THERE WAS NO DASHBOARD. "/" redirected to the underwriting queue — one
stage, usually empty, and belonging to somebody else. Every screen in
the app answered "show me this queue"; none answered "how is the book
doing, and is anything waiting on me?", which is the only question a
person has before they have picked a queue.
So: a Today page. Three blocks, in the order somebody cares about them.
What is waiting on a person, as three cards that colour only when they
have something in them. Renewals running out, worst first, capped at 30
days because a renewal six months away is not a thing to look at today.
Then every stage with its count, zeroes included — a stage quietly
receiving nothing is only visible if its zero is on screen.
It is one list call, counted in the browser. That is honest at this size
and it is the same call the queues already make. If the book outgrows it
the answer is a counts endpoint, not a bigger page.
THE WORDS WERE WRITTEN FOR WHOEVER BUILT THE WORKFLOW. "What happens
next" was followed by "these are the only activities this state allows
— but the workflow decides, server-side, whether you may". A failed list
said "this queue has no record view yet". Panels were called "Flow
details". A closed lead read "nothing runs from here". None of that is
wrong; all of it is addressed to the wrong reader.
Now: "What you can do", "Lead details", "History", "This queue could not
be loaded". Queue names say what they want — "Documents needed",
"Underwriter to review", "Payment to confirm". The sidebar group called
"Not yet" says "Not due yet", and "Running by itself" says "The agents
have it", which is the actual claim being made.
THE QUEUE HAD NINE COLUMNS AND LED WITH THE WRONG ONE. Two were internal
vocabulary: attribution status renders "clear" or "contested" and means
nothing to an operator, and the channel is background rather than
something anyone scans a queue for. Both fold into a subtitle under the
customer. That leaves six columns and puts the renewal countdown — the
number that decides whether to act today — second instead of fifth.
Also: the timeline printed a raw slug when a step came back without a
name. It is the screen this product is demonstrated on, so a
"zk-act-doc-reminder" in the middle of an otherwise readable story is
expensive. It now reads as English.
The lead file listed the current field values and nothing about how they
got there. On an app whose whole claim is that agents do the work, that
is the wrong thing to show: you could see the outcome and not whether a
machine or a person produced it.
Adds a timeline from the audit trail — every activity in order, badged by
actor, carrying the reasoning the agent actually wrote (attribution,
eligibility, call notes, recommendation rationale, the premium breakup)
rather than the raw field dump that is already below it.
Colour encodes something real: machine work is blue, people amber, the
platform grey, so a file reads at a glance for how much of it was done by
hand.
Also fixes the audit path — the bare /view/audit is not an API route and
falls through to the SPA, returning HTML with a 200 that parses as a JSON
error rather than an HTTP one. It has to be app-scoped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
