There Is No Learning Loop: How an Agent Actually Gets Better Here

The claim we are not making

The usual pitch is an agent that learns. A customer corrects it, the system extracts a lesson, indexes it, and pulls it back into future conversations. Every agent improves on its own, forever, with a confidence score attached to make it look like measurement.

ThinnestAI does not do this. There is no learning store, no automatic capture from feedback, and no process that changes how your agent behaves without you doing something.

This is a choice, and the rest of this post is what you get instead.

Why a business might want it this way

An agent that edits its own instructions is an agent whose behaviour on Thursday is not the behaviour you approved on Monday. Three problems follow, and they are the reason regulated customers ask about this first.

• You cannot answer "why did it say that". When a complaint arrives about something the agent told a customer six weeks ago, the honest answer under a learning system is that it had absorbed a lesson from an earlier conversation you never read. That is not an answer anybody can act on.
• The corrections come from strangers. The people talking to a public agent include the ones probing it. A system that treats every thumbs-down as a lesson has handed a slow steering wheel to whoever is most persistent.
• Wrong lessons are stickier than wrong prompts. A bad line in a prompt is one line, in one place, that you can find. A bad learning retrieved by similarity fires on some conversations and not others, which is much harder to notice and much harder to reproduce.

The trade is real and worth stating: manual improvement is slower. You have to look. What you get for looking is that every change to your agent has a person and a moment attached to it.

The four things that actually change behaviour

1. Its knowledge

Most wrong answers are not reasoning failures. They are the agent not having the fact. Knowledge comes from a website crawl one to three levels deep, a sitemap import, text you paste, and file uploads — txt, md, csv, and the text layer of a PDF. There is no OCR, so a scanned price list is not a knowledge source; retype it or paste it.

CSV chunking is row-aware, which matters more than it sounds: a price that gets separated from its product is worse than a missing price, because the agent will answer confidently with the wrong one. And one knowledge base serves all 37 reply languages — you do not maintain a Hindi copy of your refund policy for the agent to get out of step with.

2. Its instructions

The system prompt is where tone, boundaries and the order of questions live. Editing it is the fastest change you can make and the one with the widest blast radius, so change one thing at a time and read the next twenty conversations before changing another.

3. Which tools it may call

Every tool an agent discovers arrives switched off, and an agent stops at eight enabled tools. Withholding a tool is the strongest instruction available, because it is not an instruction at all — there is nothing to call, so the behaviour cannot occur regardless of what a stranger types into your widget. If an agent keeps doing something you do not want, check whether the thing it is doing is a tool you can simply take away.

The eight-tool cap is part of quality, not a licensing limit. Every tool schema is re-sent to the model on every turn, and models get measurably worse at choosing as the list grows. A twelve-tool agent is a less reliable agent with a longer prompt.

4. When it hands over

Escalation triggers are per agent: hand over when the agent cannot answer, hand over when the customer asks for a person or is upset, or neither. Turning one off removes the tool rather than discouraging its use. An agent that escalates in the right places is often a bigger improvement than an agent that answers more questions.

What to read before you change anything

• Transcripts. Every call has one, and it is the conversation's own messages rather than a second store — which is why a phone call and a WhatsApp thread with the same customer sit in one history. Recordings play back in the console and are deleted after 30 days on Pay As You Go, 75 on Scale.
• Call outcomes. Connect rate, average talk time, and why the rest failed: no answer, line busy, declined, not reachable. Most first campaigns are fixed by changing the calling hours, not the script — you only find that out by reading the disposition breakdown before the connect rate.
• Captured leads. Each one deep-links back to the conversation it came from. Reading the ten conversations that produced leads and the ten that nearly did is a better use of an hour than any dashboard.
• Escalations. The agent writes its reasoning for handing over into an internal note. A week of those is a list of the questions your knowledge base does not answer, in your customers' own words.
• A CSV export of the whole result set — it runs its own query rather than serialising the rows somebody scrolled to.

A routine that works

Read ten conversations a week. Not a sample the system chose for you — ten you pick, including the boring ones. Write down every question the agent could not answer. Fix the knowledge base for the ones that are missing facts. Fix the prompt for the ones where it had the fact and used it badly. Turn off a tool it keeps reaching for at the wrong moment. Change one thing per round so you can tell which change did what.

Ten failures read properly will tell you more than a composite quality score would have. That is also why there is no AI judge here marking your conversations out of ten — a number produced by one model reading another model's output moves when you change the judge's prompt, and it is very easy to mistake for evidence because it has a decimal point in it.

The one thing that is automatic — and it is not learning

The agent does recognise a returning customer. History from other channels is injected as a short digest, capped at six messages and 700 characters, including on a live phone call, and phone callers are matched to their WhatsApp and web history by number.

That is memory of a person, not a change to the agent. It makes one conversation better; it does not alter how the agent behaves for anybody else. And because caller ID is spoofable, records are linked rather than merged, and the default disclosure tier lets the agent say you have spoken before without repeating what was said until the person verifies.

Get started

Open your call log, read ten transcripts end to end, and write down every question the agent fumbled. That list is your next knowledge upload and your next prompt edit, in that order.

Build an agent free →

No credit card required • Trial includes 25 voice minutes and 200 chat replies