Voice AI agent self-improvement: daily call reviews and 290+ updates in one month

The loop in one paragraph

While the campaign ran, the loop looked like this. The agent made calls; every conversation was recorded, transcribed and translated into the team's language. Once a day we reviewed the conversations for mistakes, complaints and unanswered questions. Each mistake was logged, fixed with a rule in the script or the logic, checked on test conversations and only then shipped. A few days later we checked whether it held up on live calls. It works the same way for Polyglot clients.

Where improvements come from

Three sources:

• The daily review. A separate AI agent reads yesterday's calls and messages and builds a list: where the agent answered wrong, where the client was left without an answer, where we promised too much.
• Complaints and frustration. Any client line with irritation in it goes into a separate register. We fix the script right away and close the entry with a note on what changed.
• The owner's notes. The business owner listens to some calls himself. His lines, like “why send it to the team when we have the quotes,” become rules almost word for word.

From Sep 26 to Oct 25, 2026, the changelog shows 293 releases—290+ updates. That covers everything: the voice agent, the CRM, the chat responder, automated checks. Dozens of them are fixes to the agent's own behavior from the error log—from one phrase in the greeting to new funnel steps.

The error log

Every mistake found is its own entry:

The last row matters most. Without it, a month later you can't tell whether the rule still works or another fix quietly broke it.

Testing before deploy

Every rule is tested against a set of test scenarios: “the receptionist asks to leave a number,” “the client says 'no' over the pitch,” “they ask '¿Cómo?'” and so on. A fix ships only if the old scenarios stay green and the new one passes.

And there are rollbacks. Once we switched the agent to a “static” prompt to make calls cheaper. The first line started coming in after 3–4.5 seconds instead of two—long enough for the other person to decide nobody's on the line. We reverted to the previous version an hour and a half later.

Why a prompt needs a cap

The least obvious lesson. You want to write every fix into the script as its own story: “On September 25 a client said X—answer Y.” In 12 days without limits, the agent's prompt grew from 24,000 to 48,000 tokens. Calls got more expensive, and the agent started tripping over its own rules.

Now every agent has a size cap. A new rule only goes in as a replacement: we generalize similar cases and remove outdated, dated stories. Every edit is recorded in a version log—who, when, why, and the character count before and after.

What the client gets

• The agent improves on your own conversations, not abstract examples.
• Every change is transparent: you can see what changed and why.
• Recordings and transcripts are yours: you can hear every conversation and read it in your own language.

FAQ

Does the agent learn on its own, without people?

No. AI helps find the mistakes, but a person decides which rule to add and does the testing. Fully automatic “self-learning” in sales is risky: the agent can learn the wrong thing.

How fast does a mistake get fixed?

Simple phrase fixes—the same day, after testing on the test scenarios. Funnel logic changes—within a few days, sometimes shipped switched off at first until the owner decides.

What does “290+ updates a month” mean?

These are system releases from the changelog for Sep 26 – Oct 25, 2026: 293 versions in total—agent, CRM, chats, checks. Dozens of them are agent-behavior fixes recorded in the error log. It is not “290 prompt versions.”