Versions

Every time you publish a change to your agent in Retell, a new version goes live. SignalMD records what changed and compares production calls on the new version with calls on the version before. It calls a change better or worse only when the difference is real, and says so when there are not enough calls yet.

​

When SignalMD picks up a new version

SignalMD picks up a new version within about 10 minutes of publishing, or at once when you publish from the copilot’s Build page. It then captures the new version’s configuration, records what changed since the last captured version, updates your board, and starts a run of your release tests if testing is set up (see Test a fix). A change records steps, prompts, transitions and tools added, removed or rewritten, a new model, and changed agent settings such as voice, responsiveness and interruption sensitivity. Versions that went live before you connected still appear, because every call carries its version, but Retell serves only the live configuration, so what changed in them is unknown. New calls arrive on their own (see Connect Retell), so a comparison grows as calls arrive on the new version.

​

Where to find it

On Agent config, History lists every version with its First call, Calls and Got it done, plus Handed to a person when any call ended in a transfer. A version whose configuration was never captured is marked “no config”. Below it, each change such as v7 → v8 shows what Changed in the configuration next to its Effect on production, with a conclusion tag and a written summary. A change whose configuration was not captured on one side is tagged Calls only. Each version is also a marker on the overview board’s charts, at its first call. Hover it for the configuration changes and a before-and-after reading of that chart’s metric.

​

What is compared

The Effect table uses the overview board’s own definitions, so the two never disagree: Got it done is the headline and the others are guardrails, so a change that raises Got it done while breaking something else does not read as a clean win. Got it done and the outcome rows count calls where the caller spoke and the call has been read; Calls with a failed tool counts calls where a tool ran. Hover a cell to see its number of calls. Handed to a person is never coloured and never counts for or against a version, because for many agents reaching a person is the designed result. A row with nothing on either side is left out. Rows SignalMD has not finished reading on these calls are left out too, and named under the table after “Not shown”. Below the metrics, Issues that moved lists up to 5 issues whose share of calls changed by a real amount between the two versions. They count towards the conclusion like the guardrails.

​

Verdicts you can see

Each row gets a verdict, and the change as a whole gets a conclusion tag. A movement counts only when SignalMD finds it real and large enough to act on, so a small wobble on a busy agent, or a large swing on a handful of calls, does not read as a result.

​

On a single metric

​

On a version change

The tag next to v7 → v8 sums up the rows that have a better direction, and the issues that moved: A guardrail can carry the conclusion while Got it done is unchanged: a faster Reply time alone reads as Observed improvement, and the summary says which row decided it. What to do with each is in Compare versions.

​

Why a small change can read as unchanged

The fewer calls a version has, the larger a change has to be before SignalMD can tell it apart from chance. On a quiet agent, a modest real improvement can show as Within noise for a while; it becomes visible as calls arrive. If you expect a small effect, plan to wait, and use tests for an earlier answer.

​

What a comparison cannot tell you

• It is not a controlled experiment. Retell ships a version to all traffic at once, so the callers, the season or a system behind a tool may have changed at the same time.
• Several metrics are checked at once. Now and then one moves by chance alone. A single changed guardrail, with nothing else, is weak evidence.
• Numbers from SignalMD’s reading of each call can shift slightly. Two readings of the same calls can disagree on a few of them, so treat a borderline result with care.
• It cannot say which change moved an issue. A version bundles every change in it. For one issue, read the sentence above By version on the issue page, which waits for the agent’s grace period (7 days by default). See Did my fix work.

​

Next steps