Evalgent
Back to Blog
Voice AI Testing

Voice Agent Logs vs What Actually Happened: Where the Agent's Own Account Goes Wrong

Deepesh Jayal
14 min read
Voice Agent Logs vs What Actually Happened: Where the Agent's Own Account Goes Wrong
On this page

A caller says "Thursday afternoon." The agent says "You are booked for Thursday morning," and its log shows book_slot: success. The booking system has no record of the slot. The carrier shows a call that lasted nine seconds longer than the agent thinks.

Which record is right? The usual habit is to trust the agent's own log, because it is the most detailed one you have. That habit is the problem. The agent is a party to the call, and it wrote its own account.

This post is about voice agent logs accuracy. It shows where the agent's account drifts from what happened, which independent records can check it, and how to say "undetermined" when none can. It uses illustrative numbers only and makes no benchmark claims. Provider facts are from the docs we read, as of November 21, 2026.

The agent's own log is one witness, not the record

Two neighbouring posts assume the evidence is sound. Voice agent logs accuracy is the missing step. Our 10-minute debugging runbook tells you how to read a call. Our root cause analysis post tells you how to group failures into causes. This post asks the earlier question: is the evidence true?

The idea comes from monitoring practice. Google's SRE book defines black-box monitoring as "testing externally visible behavior as a user would see it." It defines white-box monitoring as metrics exposed by the internals of the system. An agent's log is white-box. It tells you what the agent believed. Carrier and downstream records sit closer to what actually happened outside the agent.

Neither kind is wrong by nature. The same chapter says monitoring should answer two questions: what is broken, and why. A log written by the suspect can help with why. It should not be the only witness to what.

Seven ways the agent's account drifts from reality

Each pattern comes from how logs get written. The table pairs each with the record that can check it.

PatternWhat the agent log saysHow reality can differIndependent record
Requested, not completedAction logged at the requestThe provider never finished itProvider log, delivery status
Transcript is STT outputThe caller said XThe recognizer guessed XNone among these records
Tool success, downstream rejectionTool returned successThe system of record refused itDownstream record
Call end and dispositionEnded at T, resolvedCarrier shows another duration or statusTelephony call record
ClocksEvent at time TAnother system stamped another timeProvider timestamps
Dropped or reordered eventsOrder of arrival is the order of eventsEvents arrived late or neverSequence numbers
Sampled or truncated loggingNo line means nothing happenedThe line was dropped or cutRecords versus calls count
Timeline of one illustrative booking across agent log, provider log and downstream record: the agent logs success at the request, the provider accepts it, then the downstream system rejects the booking.

Logged when requested, not when completed

Most voice agent call logs record an action when the agent issues it. The line says "sent" or "ended." It cannot say "done," because done happens later, somewhere else.

Twilio's Media Streams documentation shows the shape of this. Audio your app sends is buffered and played in the order received. To learn that playback finished, you send a mark message after the audio. Twilio sends the mark back when that audio ends. A clear message empties buffered audio and also causes marks to come back.

So "spoke the confirmation" often means "queued the confirmation." A mark proves the buffer drained or was cleared, not that the caller heard the whole sentence. Log both the send and the mark, and treat the gap as data.

The same pattern covers texts, emails and hangups. "SMS sent" usually means "SMS accepted by the API." Delivery is a later fact, as the Twilio 30034 post shows.

Voice agent transcript accuracy: the transcript is the recognizer's account

A transcript reads like a record of the call. Voice agent transcript accuracy is bounded by a speech-to-text model's output, not the audio. Deepgram's documentation describes interim results as a "preliminary transcript, may change." A final transcript covers one audio segment, and the speaker may not have paused yet. Check what your log stored: an interim, one final, or all finals joined. Deepgram advises appending each final to a buffer. A log that keeps one fragment misstates the utterance.

The deeper limit is plain. No telephony, provider or downstream record holds the caller's actual words. The only independent witness is the recording, which a person must play. This post does not cover audio review; see transcript versus audio in evaluation. Here, a claim about what the caller said is usually undetermined.

Tool "success" while the downstream system said no

A tool wrapper logs success when the call returns or no exception is raised. The downstream system may have accepted the request and rejected the booking afterwards. It may even return an error body with a success code.

OpenTelemetry's documentation says a span status of Unset "means that the operation it tracked successfully completed without an error." That describes the operation, not the business result. A clean span is silence, not proof. See silent tool failures and post-call write-back testing. The habit is the same: ask the system of record.

The call is logged as ended, the carrier disagrees

Your agent decides when the call ended from its own events: a hangup request, a closed socket, a finished session. The carrier counts its own way, in what telephony people call call detail records.

On Twilio call duration vs agent logs, the Call resource defines duration as "the length of the call in seconds" and leaves it empty for busy, failed, unanswered or ongoing calls. Its completed status means the call was answered and ended normally, with a connection made and audio transferred. It says nothing about whether the caller's goal was met.

"Completed" at the carrier and "resolved" in the agent's log are different call disposition claims. Only the first has a carrier witness.

Compare like with like. Twilio's Voice Insights call summary reports duration from initiation to end and connect_duration from answer to end. Its documentation says a partial summary is available within ten minutes, and a complete one may take up to half an hour. An early audit reads incomplete data.

Clocks, and events that arrive out of order

Every system stamps events with its own clock. OpenTelemetry's log data model separates Timestamp, when the event occurred by the origin clock, from ObservedTimestamp, when the collection system observed it. The first may be missing if the source time is unknown. The two can differ.

Arrival order is a second trap. Twilio's status callbacks carry a SequenceNumber that starts at 0. The Call resource page says there is "no guarantee they will arrive in the same order." A handler that keeps the last callback it received can end a call in the wrong state.

Two illustrative panels: callbacks arriving as sequence 0, 1, 3, 2, where last arrival shows answered but sorting shows completed, and a booking system receiving a request 3.3 seconds before sending.

The fixes are cheap. Sort by sequence number, not arrival. Store both timestamps where a system gives you both. Measure the offset on a known event before trusting any ordering across systems. A record that precedes its own request signals a clock offset, not a time machine.

Logging that is sampled, truncated or dropped

Absence of a log line proves nothing when the pipeline may have dropped it, and voice agent logs accuracy suffers most there. OpenTelemetry's sampling guide says head sampling decides at the start of a trace. Its stated downside is that you cannot make a decision based on data in the entire trace, so you cannot ensure that traces containing errors are kept. Tail sampling decides after most spans are available.

Truncation is similar. The OpenTelemetry common specification has an attribute value length limit, default infinity, past which SDKs must truncate strings. Its attribute count limit, default 128, discards extra attributes. A long tool payload can be cut with no error.

The test is a count. Compare records to calls, since voice agent logs accuracy includes completeness. If the carrier shows 1,000 calls and your traces hold 940, 60 calls have no agent account. Those calls are not clean. They are blind. Review sampling is a separate question; see sampling live calls.

What each independent source can and cannot witness

An independent source is one written by a different system, triggered by a different event, with no shared code path to the agent's log. A record your own agent writes to the database is the agent's log in another place. It is not independent.

SourceWhat it can witnessWhat it cannot witnessWatch for
Carrier or telephony call record (Twilio Call resource, Voice Insights summary)Whether the call connected, its status, duration in seconds, timestampsWhat was said, whether the task succeededEmpty duration on unanswered calls; partial summaries early
Status callbacksState changes as the carrier saw themA guaranteed orderSort by SequenceNumber; late or missing callbacks
STT provider log or dashboardWhat the recognizer returned and whenWhat the caller saidInterim versus final text
Messaging or email providerWhether a message was accepted, delivered or failedThat the caller read itAccepted is not delivered
Downstream system of recordWhether the booking, ticket or update existsWhy the agent said what it saidIts own clock and validation rules

One call, reconciled

This reconciliation table is illustrative. The call, the numbers and the systems are invented to show the method.

An agent reschedules deliveries. For one call it logged the following. The call ended at 142 seconds. book_slot returned success. An SMS confirmation was sent. The outcome was "resolved." The transcript reads "Thursday afternoon." The tool request left at 14:03:12.4 by the agent clock.

ClaimAgent's accountIndependent recordVerdict
Call was answeredSession startedTwilio status completed, duration 151 s (non-empty)Confirmed
Call lengthEnded at 142 sTwilio duration 151 s, a 9 s differenceContradicted
Booking madebook_slot successBooking system: write rejected, slot already heldContradicted
Text confirmationSMS sentSMS provider log: delivery failedContradicted
Outcome "resolved"ResolvedCarrier says completed, which does not cover resolution; booking system shows no bookingContradicted
Caller's wordsTranscript says "Thursday afternoon"No telephony, provider or downstream recordUndetermined
Order of eventsTool request sent 14:03:12.4Booking system received 14:03:09.1Undetermined: clock offset of at least 3.3 s

The arithmetic: 151 minus 142 is 9 seconds. 14:03:12.4 minus 14:03:09.1 is 3.3 seconds. Of seven claims, one is confirmed, four are contradicted and two are undetermined.

The nine-second gap has no known cause. The agent may have logged a hangup request while the carrier leg stayed open. Without playback marks or a carrier hangup event, that stays a hypothesis.

The booking row is the one that matters. The caller left believing something that the record says is false, and the agent's log hid it.

When no independent source exists, say undetermined

Some claims have no witness among your records. What the caller said is the standing example. Pretending otherwise turns a gap into a false verdict.

Use three verdicts, and only three.

  • Confirmed: an independent record agrees with the claim.
  • Contradicted: an independent record disagrees. Say which.
  • Undetermined: no independent record covers the claim, or the records conflict and nothing breaks the tie.
Decision flow for each agent log claim: if an independent record covers it, compare to confirm or contradict; if none exists, mark it undetermined and name the missing source.

An undetermined verdict should name the missing source. That list is your coverage gap, and it is a to-do list for logging. Keep it as a short record per claim.

CLAIM LEDGER ENTRY
Claim:      Caller asked for Thursday afternoon
Agent says: Transcript turn 6
Source:     none among telephony, provider or downstream records
Verdict:    UNDETERMINED
Missing:    Playable recording reviewed by a person, or a scripted test caller
Next step:  Add the call to the review queue; do not change the prompt on this evidence

Evalgent Observe, which is launching, follows the same rule at the product level. A verdict needs evidence from the data you supply. Where that data is blind, it says "undetermined," and the coverage gap is shown, not hidden. Observe works from the traces and call data you send it, through OpenTelemetry plus an upload path for call details.

In a test, the problem shrinks. You scripted the caller, so you know what was said. Evalgent Reviews show the exact tool payloads of a test call, which makes the "requested versus completed" check concrete before launch.

How to reconcile voice agent logs against independent records

Run this routine after any complaint, and on a schedule for a small random set.

1. List the claims — Write down what the agent's log asserts: call length, answered, each tool result, each message sent, outcome. One line per claim.

2. Map each claim to a witness — For every claim, name the independent record that could confirm it. Use carrier records, provider logs and downstream records. If you find none, write "none" now.

3. Pull the records late enough — Wait for partial data to complete. Twilio's summary can take up to half an hour, and its Call resource warns a recording may not be accessible when the status callback fires. Retry before concluding anything is missing.

4. Normalize before comparing — Convert to one unit and one clock. Compare duration with duration, and connect time with connect time. Sort events by sequence number, not arrival. Measure the clock offset on a known event.

5. Assign a verdict per claim — Mark confirmed, contradicted or undetermined. For each contradiction, name the record that disagrees. For each undetermined, name the missing source.

6. Count and fix the blind spots — Compare record counts to calls. Then change logging so the next audit finds fewer undetermined rows.

Change your logging so the next audit is easier

Most contradictions are cheap to prevent if the log keeps the right facts. What to log on every call has the full field list. Five habits matter here.

  • Log request and result as two events. "Requested" at send. "Completed" or "failed" when the other side answers.
  • Store provider identifiers. Keep the carrier call ID, message ID and downstream record ID, so you can fetch the independent record. See call ID correlation and tracing a call end to end.
  • Keep raw and final text. Store interim and final transcripts with their flags.
  • Keep both times. Store the origin time and the time your pipeline saw the event.
  • Log the log's own gaps. Record sampling decisions, truncation and dropped counts, so absence can be told from silence.

With these in place, why voice agents fail in production is easier to read, because each layer has a witness. For the wider frame, see what is voice agent observability.

Where Evalgent fits

Evalgent is a testing, evaluation and monitoring platform for voice agents. Its production call monitoring flags problems such as tool-call failures and latency spikes, with fix recommendations. Observe is aimed at the explanation step: it traces each failure to the step or handoff that caused it, and clusters failures into root causes. Evalgent does not sit in the media path, so nothing touches live audio or slows a call.

You can run the routine above with a spreadsheet and the carrier's records. To talk through evidence-backed verdicts for your own agent, book a demo.

Frequently asked questions

Are voice agent logs reliable?

Partly. They are reliable about what the agent decided and requested. They are weak about what completed, what the caller said and what the carrier recorded. Treat the log as one witness. Check completion claims against provider and downstream records, and mark claims with no independent source as undetermined.

How do I verify voice agent actions?

Pair each action with an independent record. A booking needs a record in the booking system. A text needs a delivery status from the messaging provider. A call needs a carrier call record. Compare after the action, retry for records that arrive late, and keep the provider ID on the agent's log.

Is the voice agent transcript accurate?

A transcript is the speech recognizer's output, not a recording of the caller's words. An interim transcript can change, and a final transcript is the recognizer's best account of a segment. No telephony or downstream record holds the actual words. Claims about what was said are usually undetermined without a person reviewing audio.

Why is Twilio call duration different from my agent's log?

They measure different things. Twilio's duration is the length of the call in seconds as the carrier counts it. Your agent may log the hangup request, the socket close or the session end. Voice Insights also separates total duration from connect duration. Compare matching measures before calling a difference an error.

What does Twilio's completed status mean for the agent's outcome?

It means the call was answered and ended normally, with a connection made and audio transferred. It does not say the caller's goal was met. "Completed" at the carrier and "resolved" in the agent's log are separate claims. Check resolution against the downstream system of record.

Why does a tool call show success but the booking failed?

Tool success usually means the request returned without an exception. The downstream system may have rejected the booking afterwards, or returned an error inside a success response. In OpenTelemetry, an unset span status means only that the operation finished without an error. Verify the result in the system of record.

Can status callbacks arrive out of order?

Yes. Twilio's Call resource documentation says there is no guarantee that status callbacks arrive in the order they were sent. Each carries a sequence number starting at 0. Sort by that number, not arrival time, so a late "answered" callback does not overwrite "completed."

How do I handle claims I cannot verify?

Mark them undetermined and name the source that is missing. Do not guess a verdict from the agent's own log. Add the missing source to your logging plan, such as playback confirmations or a scripted test caller. Track the share of undetermined claims as a measure of your blind spots.

The bottom line

The agent's log shows what the agent decided and asked for, not what the carrier, the provider and the downstream system recorded. Check each claim against an independent record, and write "undetermined" when none exists.

Related Articles