Agent Performance

Agent Performance compares your AI agents on the same measures over the same date range. It answers questions like "which agent hands off the most conversations?" and "did last week's prompt change help?". Open it from Insights → Agent Performance.

It covers conversations an AI agent handled and finished. A conversation appears once it ends. Use Supervisor for what is happening now.

Every number on the agent's page opens the conversations it counts. See Drill-downs.

The agent list

The page has eight tiles for all agents together, then a table with one row per agent. Each tile compares against the prior period, the range of equal length immediately before the one you selected. The tiles are Conversations, Containment, Resolution, Negative sentiment, Transfer rate, Take-over, Avg handle time, and Repeat in 7 days.

Agents are sorted by conversation count. Click a column heading to sort by it. Click a tile to rank the agents by that measure instead: the agents with the lowest containment or resolution come first, and for every other measure the highest value comes first. Agents with no value for the measure come last.

A rate that trails the organization's by 10 percentage points or more is flagged with a warning icon, once the rate rests on at least 30 conversations. For resolution and negative sentiment the 30 are judged conversations. For repeat in 7 days they are the conversations old enough to check. Click a flagged value to open that agent's page with the measure selected.

The filters are the same as on the other Insights pages: a date range (UTC) and, when your account has more than one channel in range, a Channel filter. See Insights & Reporting for channels and for why dates are UTC. The longest range you can select is 90 days.

An agent that handled conversations in the prior period but none in this one is still listed, with zero conversations, so a drop to zero is visible.

Click an agent's name, or anywhere on its row, to open its detail page.

What each number means

Every rate is a count divided by a stated number of conversations. Hover over a value to see the counts behind it (for example, "42 of 60").

An agent that handles the same conversation twice, for example when a workflow hands the caller back to it, counts twice in Conversations: once for each time it answered.

A value shows — when its denominator is zero. A dash means there was nothing to measure. It does not mean 0%.

Totals across agents add up the counts and then divide. A two-conversation agent therefore does not weigh as much as a two-thousand-conversation agent.

Containment counts a hang-up as contained

A caller who hangs up is contained, because the agent did not hand the conversation to anyone. A caller who hangs up in frustration still counts. Read containment next to Resolution and Negative sentiment.

Two other endings are contained too:

• The agent hands the caller back to a workflow. The workflow continues the call, so no person took it over.
• The agent switches the caller to another AI agent. This is not a transfer and does not count toward the transfer rate.

A call that ends before the agent could answer, usually because the agent could not be reached, counts as a fault, so it is not contained. The end-reason breakdown shows it as Disconnected.

Webchat conversations

Chat agents report no transfers, so their transfer rate is 0%. A chat ends only when the agent ends it or the conversation reaches its turn limit. A chat the visitor leaves without an ending never finishes, so it is not counted. Together these make containment for a chat agent read higher than for a voice agent, and abandoned chats do not appear.

Resolution and negative sentiment are judged by AI

An AI model judges these two after each call, from the conversation's post-call analysis. Wherever they appear, a sparkle icon marks them as AI-judged. Hover over it to read "Judged by AI after each call". The mark is on the tiles, the table headings, the chart title, the sheet headlines, and each conversation row that shows a verdict.

Resolution requires the Call outcome built-in field to be on for the agent. Negative sentiment requires Caller sentiment. A conversation with no analysis, or with an unknown result, is left out of the denominator.

The tile footer shows how much of the data the rate rests on, for example "AI-judged 38/60", the judged conversations over all conversations. A high resolution rate on a small share of conversations is weak evidence. Turn on the Call outcome field and write a clear definition of success to raise the share.

The analysis does not store a reason for either verdict. A conversation row shows the verdict and opens the conversation on its Analysis tab, where the call summary is. Voice calls only have a summary.

Repeat in 7 days

A conversation counts as a repeat when the same caller contacts your organization again within 7 days after it, through any agent or workflow. The follow-up matches when it has the same contact record or the same phone number as the measured conversation. A conversation with neither is not counted either way.

Calls your organization placed (outbound calls) are left out on both sides: an outbound conversation is not checked for a repeat, and an outbound call to the same caller is not a follow-up.

Only conversations that started at least 7 days ago can be checked, since a newer one has not had time to be followed up. Conversations from the past 7 days are therefore left out of both sides of the rate, whatever range you select. A range that ended a month ago excludes nothing, and a range that ends within the past week can show —. The follow-up can be on any channel. The Channel filter applies to the conversations being measured, not to the follow-up.

First-agent attribution

A conversation is credited to the agent that answered it. If that agent switches the caller to a second AI agent, the second agent has no conversation of its own, and the whole conversation counts for the first agent. The tools the second agent runs count for the first agent too, so they appear in the first agent's Tool errors.

A switch to another AI agent is not a transfer. The detail page counts it separately, in a line under How conversations ended.

The agent detail page

The detail page shows one agent over the same range and filters. From the top, it has:

• The header. The agent's name, a badge for each channel you filtered by, and a badge such as v2.1 live for the agent's published version. Open Agent opens the agent's editor. View Coaching opens the agent's page under Coaching, where Tandem recommends changes to the agent. Back to Agent Performance returns to the list and keeps the range and channel.
• Eight stat tiles. The same eight measures as the list, for this agent. The tiles are also the metric selector. A tile flagged with a warning icon trails the organization.
• The metric panel. A chart, with Changes in range under it, then What moved it.
• Where conversations go. How conversations ended, where they transferred, and which tools failed.

With no conversations in the range, the page shows "No conversations for this agent in this range" and Changes in range, which still lists any change to the agent.

The Overview and a focused metric

The panel opens on the Overview: one chart with all six rates as lines on a percent axis (Containment, Resolution, Negative sentiment, Transfer rate, Take-over, and Repeat in 7 days) and daily conversation volume as bars on a second axis. Each tile's color is its line's color on the chart. Avg handle time is in seconds, so it is not on the Overview.

Click a tile to focus the panel on that metric. The chart then shows that metric alone, a gray line marks the organization's current value (for example "Org 62.0%"), and a sentence under the title states the change from the prior period, for example "Transfer rate rose 3.1 pts, from 10.9% to 14.0%." Rates change in points. Handle time changes in m:ss. Click the focused tile again, or Show all, to return to the Overview. The metric is part of the page address, so a link opens on the same metric.

Chart and Table switch between the chart and a table with one row per day. Days are UTC. A day with no conversations leaves a gap in the chart.

Changes in range

Changes in range lists what changed on the agent between the start and end of the range, newest first, up to 200 entries. It sits under the chart, whose vertical lines mark the same changes. Each entry shows the UTC date, what happened, and who did it when the name is available.

• A version publish reads "Version 2.1 published", followed by the configuration sections that differ from the previous published version, for example "Prompt, Greeting changed". The common sections are Prompt, Greeting, Voice & models, Tools, Knowledge, Post-call analysis, Keypad input, Voice activity detection, Interruptions, Idle handling, and Data retention. Any other section shows its setting name in words, such as Turn detection or Data collection. A publish that changed none shows "No configuration changes". The version's description appears under the entry.
• A rollback is a publish of a lower version than the one it replaced. It reads "Rolled back to version 2.0 from 2.1".
• A coaching change reads "Coaching applied:" and the recommendation's title. It is listed when you mark a Coaching recommendation as applied.
• An older version reads "Version 1.3" and "Published (recorded before change history began)". Tandem rebuilds these from version history, so the date and description are known and the changed sections are not. A note below the list says so. Rollbacks before change history began show as ordinary publishes.

Versions created when versioning was first introduced for an agent are not listed, because nothing changed on that day.

The chart draws one vertical line for each UTC day with a change. A publish is labeled with its version, such as v2.1, a rollback with an arrow, such as v2.0 ↩, and a coaching change with Coaching. Several changes on one day share a line. A change outside the chart's days draws no line. If the history fails to load, the section shows "Couldn't load this agent's change history."

Click an entry, or a marker, to open its change sheet.

What moved it

What moved it breaks the change in the focused metric into the segments that moved it. On the Overview it explains one metric, and the heading names it, for example "What moved containment". The Overview picks the rate with the largest change in points. It picks Avg handle time instead when handle time moved at least 5% and its change in percent is larger. It picks Resolution only when both periods have at least 30 judged conversations.

The panel needs conversations in the range and an earlier period to compare with. Without an earlier period it shows "No earlier period to compare with." When both periods have data but no segment splits the change, for example a voice agent that took every call directly with no version split, it shows the sentence and a note that no segment split the change.

It explains Containment, Resolution, Negative sentiment, Transfer rate, Take-over, and Avg handle time. Conversations shows Volume by segment instead: how many conversations each segment had, this period and the prior one, and each segment's share. Repeat in 7 days has a chart and a sheet but no breakdown: the panel says "No breakdown is available for repeat contacts."

A contribution is the part of the change that one segment accounts for. Each segment is measured against the agent's overall value, the average of its prior and current values. A segment contributes when its own rate moved, or when its share of conversations moved while its rate sat above or below the overall value. A growing segment with a lower rate than the overall value pulls the metric down, so its contribution is negative. Two segments with the same rate as the agent overall contribute nothing, however their shares shift. The contributions of all segments in a dimension add up to the change.

A contribution shows where a change came from and says nothing about why. A segment can move the number because something else changed at the same time, such as the callers or the campaign.

Top contributors lists the five largest contributions across all dimensions. Each row shows the segment, its dimension, its movement from the prior to the current value, and its contribution. Other is left out of this list. Below it, a tab for each dimension shows a table with the columns Segment, Conversations (in the current period), Prior → Current, Contribution, and Driven by.

Driven by says which part of the contribution is larger:

• Performance: the segment's own rate moved.
• More of this traffic or Less of this traffic: the segment's share of conversations moved. This is called mix.

The dimensions are:

End reason is left out for Containment and Transfer rate because the end reason is part of how both are counted, so splitting by it would repeat the metric. For analysis fields, the page shows at most 5, the ones present on the most conversations, and only fields with 12 or fewer distinct values across both periods. A conversation with no value for a field falls in Not set. A dimension is shown only when it has at least two segments besides Other. With one, that segment would carry the whole change, which the sentence already states. Unknown in Version covers conversations with no recorded version, such as those from before versions were recorded.

A segment with fewer than 10 conversations in both periods folds into Other, so a handful of calls does not produce a large swing. A segment that has conversations in only one period has no rate in the other, and its whole contribution counts as mix. When every conversation moved from one version to the next, each of the two versions accounts for half of the change.

Transfer destination works differently. A row's Prior → Current is the destination's share of all conversations, for example "8.0% → 12.5% of calls", and its contribution is the change in that share. Conversations that were not transferred are not listed.

Click a contributor, or a segment row, to open its segment sheet.

Where conversations go

Where conversations go has up to three cards. A card is left out when it has no rows.

• How conversations ended counts conversations for each end reason, with each one's share. Handed back to workflow means the agent passed the caller to a workflow. Switched to another agent means the call went on with another AI agent, however it ended after that. Disconnected (session ended) and Disconnected (connection lost) mean the call ended before the agent could answer, and both count as faults. The other labels are Completed, Caller hung up, Transferred, Fault, Taken over, and Turn limit reached. An end reason Tandem does not list shows as stored, and one with no value shows as Unknown. When the agent switched any conversation to another AI agent in the range, a line under the heading says so, for example "3 conversations were handed to another agent." A conversation counts once however many times the agent switched in it.
• Transfers shows attempts, connected transfers, and success rate for each destination, when the agent attempted any. A transfer with no destination name shows as Unlabeled. Switches to another AI agent are not listed.
• Tool errors lists the tools with at least one failed call, with calls, errors, and error rate. It lists up to 10 tools, including tools run by an agent this one switched the caller to. What counts as a failed call is under Tool errors.

Click a row in any of the three to open its sheet.

Drill-downs

Each number on the agent's page opens a sheet beside the page with the conversations behind it. The page stays where it is, and the sheet's address is in the browser's address bar, so you can paste the link into a ticket and it opens the same sheet. From 1800 pixels of window width the sheet sits beside the page. On a narrower window it covers part of the page. Close it with the close button or the Escape key. The browser's Back button also closes it.

What each click opens

The button on the focused metric reads, for example, "View 14 Uncontained". The count is the numerator of that metric's rate: Uncontained is the conversations not contained, Unresolved and Negative are the judged ones, and Timed Conversations is the sample behind the average handle time. A marker shared by several changes opens the earliest of that day. The other changes that day are listed on the day's sheet.

Every sheet has the same parts. The headline gives an exact count, with what it is a share of underneath. A breakdown comes next, then the matching conversations, with Load More for further pages. Each conversation shows when it ran, its channel, and its duration, with a line of evidence for what the sheet is about. The footer has Open these conversations in Call Logs, which opens Call Logs with the same conversations. A sheet that carries no conversation filter reads "Open all this agent's calls in Call Logs".

What each sheet shows

• Metric sheet. The headline is the count behind the tile, for example "14 not contained" with "of 60 conversations (76.7% contained)". Containment adds Why they weren't contained: the transferred, taken over, and faulted conversations. Transfer rate adds a By destination table. Conversations adds By channel and By started by tables. Resolution and Negative sentiment carry the AI mark and list the conversations judged unresolved or negative. Take-over lists the conversations a supervisor took over. Repeat lists conversations that were followed by another, with a link to each follow-up conversation. Avg handle time lists every conversation in the range, longest first, and says so, because the average rests on the timed ones only.
• Day sheet. The conversation count for the day and a table of that day's values against the range's values, for the focused metric or for each rate on the Overview. It lists any change that landed that day, and the day's conversations.
• Change sheet. For a version publish: when it was published, by whom, the version it replaced, its release note, what changed in the configuration, a before and after, and the conversations that version handled. For a coaching change: the recommendation, its reason, the change it made, the conversations it cites as evidence with a quote from each, and a before and after. Open Version in Agent and Open Coaching go to the version and to Coaching. A change from outside the selected range says so, and the sheet asks you to widen the range before it compares anything.
• Segment sheet. The segment's prior and current value, its contribution, and what it is Driven by, then its conversations. A segment made of Other lists no conversations when it folds too many small segments for one request, and says so. Open one of the segments from the table instead. Conversations with no recorded channel cannot be listed by channel.
• End reason sheet. The count and share of that end reason, and its conversations. For Caller hung up, each row carries the AI resolution verdict, because a hang-up counts as contained and the verdict says whether the caller was helped.
• Destination sheet. Attempts, connected, and failed transfers. A Failed only switch limits the list, and each row is marked Connected or Failed. Tandem does not record why a transfer failed, so the sheet shows no cause.
• Tool sheet. See Tool errors.

Before and after a change

A change sheet shows the focused metric, or each rate on the Overview, for the seven days before the change and the seven days after it. Each side shows its n, the number of conversations the value rests on. If the week after has not finished, the sheet says its numbers are partial.

The comparison makes no causal claim. When at least one row differs between the two sides, the sheet says "This change coincided with the difference. Other events in the same two weeks could explain it." A different mix of callers or a campaign in the same fortnight moves the numbers too. When every row is equal, it says "No difference in the seven days on either side." Read a small n as weak evidence.

Opening a conversation at the moment that explains it

Click a conversation in a sheet to open it. The conversation page opens at the moment the sheet is about. A tool, transfer, or agent-switch row opens at the call with that tool name and opens it. On a chat the page picks the call nearest in time to the one the sheet listed. On a voice call, whose turns have no time, it picks the first failed call of that name, or the first call of that name when none failed. A take-over row opens at the line that marks the take-over, "X took over the call", and highlights it. A row with no such moment opens at the top. A resolution or sentiment row opens the Analysis tab instead. The Back to Agent Performance button returns to the same sheet.

The moment needs the conversation's transcript, which Tandem keeps for 90 days and only when the agent stores transcripts. See Transcript retention. Without one, the page opens at the top with a note: "This conversation's transcript isn't available, so the page opens at the top." If the moment cannot be matched, the note reads "Couldn't find that moment in the transcript, so the page opens at the top." Neither is an error.

Tool errors

The Tool errors card and the tool sheet count failed tool calls. A failure is one of these:

• A timeout, including an HTTP 408 or 504 answer.
• A server error, an HTTP 5xx answer.
• An exception, a tool that threw an error.
• Invalid parameters, a call the tool refused because its arguments failed validation.

An integration's 4xx answer returned as data, for example a 404 "not found", is recorded but is not counted as a failure. The agent can handle a "no" from the other system, so it does not raise the error rate. A tool that reports its own error is a failure whatever the HTTP status, so a 4xx exchange with an error returned counts. A tool whose only errors are 4xx answers returned as data is not listed in the card.

Each failed call, and each 4xx answer, records a kind, and the tool sheet's By error kind table groups calls by it:

A 4xx call that returned an error has its own row, Returned 4xx with an error, and counts as a failure. The row Returned 4xx (not counted as failure) holds only the 4xx answers that did not fail.

Conversations recorded before error kinds existed read as Unknown: their failed calls have no kind.

The tool sheet opens with "N of M calls failed" and the number of conversations they hit. It then has the by-kind table, the Failing conversations list, Failures by day as small bars with a table behind Show as table, and a By version table. The Include 4xx answers switch appears when the tool had any 4xx answers that did not fail. It adds a second list headed "Conversations where the tool returned 4xx". A conversation whose 4xx call failed appears in both lists. Open Tool in Agent opens the agent's editor, since there is no link to a single tool's settings. Each failing conversation opens at the failed call. The sheet itself shows no error message text, since only the kind is recorded. Error messages stay in the conversation's transcript.

Troubleshooting

• Resolution is —. No conversation in the range has an outcome of success or failure. Turn on Call outcome in the agent's post-call analysis.
• Negative sentiment is —. No conversation in the range has a sentiment. Turn on Caller sentiment in the agent's post-call analysis.
• Repeat in 7 days is —. No conversation in the range is both old enough and tied to a contact or phone number.
• A change is missing from Changes in range. Only publishes and applied coaching changes are listed. Edits to a draft that were never published do not appear.
• A tool I expect is missing from Tool errors. The card lists tools with at least one failed call. A tool that only returned 4xx answers has none.
• A conversation opens at the top. Its transcript was not stored, is older than 90 days, or does not contain the moment. The note at the top of the page says which.
• An agent is missing. The list shows agents that handled a finished conversation in the selected range and channels, in this period or the prior one.
• The page shows an error after you open a link. The date picker offers presets of up to 90 days, so a longer range comes only from a URL that was edited or pasted. Pick a preset from the date range menu.

Next steps

• Post-call analysis
• Coaching
• Insights & Reporting