Chat tool telemetry

Find slow, failing, and repetitive agent tools without exposing prompts, arguments, or outputs.

PeriodToolTeam

Detailed production timing active

495 of 570 saved tool calls include execution duration.

86.8% timing coverage

Tool calls

-29% vs previous 7d4 saved production turns

p50 latency

43 ms

Timing not instrumented5 timed calls

p95 latency

615 ms

Timing not instrumented5 timed calls

Tool failures

80.0%

-20.0 pts vs previous 7d4 excluding NoResults

No result

0.0%

No change vs previous 7d0 product outcomes

Calls / turn

1.3

+7% vs previous 7d100.0% of turns at 5+ calls

What changed

Material regressions ranked by severity and affected call volume.

Current 7 days compared with the previous equal window

Stable against baseline

No material tool regressions detected

Tools with enough activity were checked for higher failure rates, available p95 latency regressions, more no-results, and increased loop pressure.

Review the tool comparison for absolute hotspots that may still need a product threshold.

Investigate

Regression checks require at least 3 calls in both windows. Latency is flagged at +20%; failure and no-result rates are flagged at +1 and +3 points respectively.

Execution volume

Successful calls, no-results, and failures combined.

Calls

Aug 1Aug 4Aug 7

Payload pressure

How often results exceed the 20 KB inline cap.

Unavailable

handle rows / all calls

The handle store could not be read with the current credentials.

Tool comparison

Each change is measured against the previous equal window; latency appears only where detailed timing exists.

Sorted by current p95 latency

Tool	Calls	p50	p95	Failures	No result
load_office_document	5-29% vs prior	43 msNo prior timing	615 msNo prior timing	80.0%-20.0 pts vs prior	0.0%No change vs prior

Failure codes

Operational failures only; NoResults is excluded.

Unknown4

Interpretation notes

The data has known edges that change what a chart means.

Blocked loops are invisible: Rate-limited and identical-argument calls are rejected before telemetry fires.
NoResults is not downtime: It remains a failed tool outcome, but is separated from runtime reliability here.
Preview volume lives elsewhere: Preview and threadless chats reach PostHog but are intentionally absent from Postgres.
Latency is partially instrumented: Saved tool parts provide production outcomes and volume; duration appears only on calls with detailed telemetry.
Partial turns are retained: They remain in volume counts but are excluded from failures and latency percentiles.
Comparisons use equal windows: Every delta compares the selected period with the immediately preceding period under the same tool and team filters.

Recent sanitized calls

Arguments, outputs, user IDs, and message content are never rendered.

Privacy safe

load_office_documentHJC615 msAug 5, 8:40 PM UTC

Outcome

Success

Thread

d02bb273…

Tool call

call_lHv…

Turn sequence

2 of 6

Turn state

Completed

load_office_documentHJC43 msAug 5, 10:00 AM UTC

Outcome

Unknown

Thread

044f7884…

Tool call

call_T7S…

Turn sequence

9 of 18

Turn state

Completed

load_office_documentHJC61 msAug 3, 3:46 PM UTC

Outcome

Unknown

Thread

f5d38ce1…

Tool call

call_Bbb…

Turn sequence

22 of 28

Turn state

Completed

load_office_documentHJC40 msAug 3, 3:22 PM UTC

Outcome

Unknown

Thread

490ed0f1…

Tool call

call_uKZ…

Turn sequence

4 of 9

Turn state

Completed

load_office_documentHJC41 msAug 3, 3:22 PM UTC

Outcome

Unknown

Thread

490ed0f1…

Tool call

call_07q…

Turn sequence

8 of 9

Turn state

Completed