Every AI agent has a list of things it does not know. Your return window for sale items. Whether the Austin branch does bodywork. What to say when someone asks for a discount. Which order number format your system expects. The list exists; it is scattered across two hundred conversations where the agent said let me connect you to a colleague, and the only way to read it has been to read all two hundred. Agent Auto-Improvement reads them for you, and then it asks.
Nobody Wants to Listen to Every Call or Read Every Chat
Everything the agent missed is already recorded. Nobody has time to go through it.
Your AI agent talks to customers on WhatsApp, on email, on calls, in web chat, on Instagram and Messenger. Every one of those conversations is saved: the chats as transcripts, the emails as threads, the calls as recordings. Somewhere in there is exactly what the agent does not know yet.
But nobody reads hundreds of WhatsApp chats. Nobody opens every email thread. Nobody listens to call recordings one by one to hear the moment the agent said let me connect you to a colleague. Each of those moments is small and reasonable on its own. Together they are a list of the ten things your agent most needs to learn, and nobody has the afternoon to put it together.
Agent Auto-Improvement puts it together for you. It reads and listens to every conversation, on every channel, and turns what the agent missed into one page on each agent, headed What your AI agent has learned it needs from you: a to-do list, short, ranked, and built entirely from what actually happened.
Why a Request Is Created
Only when real conversations show the agent is missing something.
- The agent's conversations on every channel are read. Calls, WhatsApp, email, web chat, the lot. Conversations that evaluation flagged as weak are read first, ones it scored as fine are skipped. Personal details are masked before anything is analysed.
- What it could not answer or do is collected as evidence. A question it handed over. A lookup it could not make because nothing is connected. An offer it would not make because it had no permission. An intent that failed.
- Repeated gaps become requests. Identical questions are grouped, then similar ones are clustered. A gap needs to appear in at least three conversations before it becomes a request, and the conversations travel with it. There is no you have a table, why not use it? Every request is caused by something a customer said.
- You answer once, or let it improve itself. Answer a request yourself, ignore it, or switch on automatic improvement and let the agent apply the answer on its own. Either way, the answer becomes knowledge, an intent, a setting or a guideline, and the agent uses it from the next conversation.
Four requests, thirty-five conversations behind them. Each one opens to the transcripts that caused it.
Five Kinds of Request
The kind decides what your answer turns into.
Alongside the requests, the page keeps a short list of suggestions from your workspace: a table that looks like it answers a question the agent keeps getting, a knowledge source another agent already uses, a website to read when the agent has no knowledge at all. Those are one click to connect. And a panel headed What this agent knows lets you paste text, give it a website, or upload a document without waiting to be asked.
Answering: What Your Answer Becomes
Write it, attach it, connect it, grant it, or say no.
Open a request and the answer bar offers what fits its kind. For knowledge: Write an answer, Attach a document, Add to a table. For a data connection: Intent Builder, with the request's own contract already filled in — when the agent should call it, which parameters it needs, what comes back. For a permission: Grant with the limits, or Skip. For a behaviour: the guideline, which you can mark mandatory.
A permission with its limits. Grant it and the agent may offer the coupon from the next chat; skip it and it keeps declining politely.
Whatever you choose, the request moves to Done with a line that says what it became — Guideline added, Knowledge added, Intent created — and a link to the version it created. Ignore and Don't ask again exist too, because sometimes the right answer to customers keep asking for X is that you do not do X.
Fix Requests: Things That Measurably Failed
No interpretation. Something ran, and it did not work.
Most requests come from reading what the agent said. Fix requests come from what the platform recorded: an intent that ran six times without a parameter it needed, a connected system that returned an error, a booking that failed, a call transfer that did not connect, a procedure step that points at an intent somebody deleted last week. No language model is involved in finding them. They are counted, and the count is the title.
Off, Prepare Answers, Apply Automatically
How far the platform may go before you look.
Automation is a single setting on the page with three positions.
- Off. The agent's conversations are not read and no requests are created. This is the starting position.
- Prepare answers. Conversations are read, requests are created, and for each one the platform drafts an answer from what it can already see — your team's replies in the same conversations, tables and sources in the workspace. You approve with one click, or edit first. Anything it cannot draft stays open for you.
- Apply automatically. Drafted answers for knowledge, read-only lookups and guidelines are applied without approval. Each one creates a version marked Automatic improvement, with the request as the reason, and the workspace owner is notified. Permissions always wait for a person. The agent never grants itself the right to offer a discount.
The middle position is where most teams will live. It turns a to-do list into an approval list, which is a different amount of work.
Did It Help?
Every answer reports back.
An answered request does not disappear. It shows Used in 23 conversations since Sep 14, and the count keeps climbing as the agent uses what you gave it. At the top of the page, two numbers for the month: what the agent learned, and how many conversations were helped by it. If a question keeps coming back after you answered it, the request says so — Still asked after your answer — because that usually means the answer was there and the phrasing was not.
Try this: switch one agent to Prepare answers, wait a day, and read the requests before you read anything else about that agent. They are the transcript review you were not going to do, already done.
What It Will Never Do
Three lines the loop does not cross.
It never edits your prompt
Behaviour answers become guidelines, which the platform folds into the generated prompt. Your prompt text stays yours. If the agent runs on a custom prompt, the page warns you that guidelines are not read until the rule is added there.
It never asks without evidence
A request needs at least three conversations behind it, and they are attached. The agent does not suggest things because they would be nice. It asks because a customer did.
It never grants itself a permission
Coupons, proactive messages, redirects, out-of-hours handoff: whatever the automation setting, these wait for a person.
It never changes anything you cannot undo
Every applied answer is a version. The two-column view shows exactly what it did, and Restore puts it back.
Frequently Asked Questions
Where do the requests come from?
From the agent’s own conversations on every channel. Conversations that evaluation flagged are read first; what the agent could not answer or could not do is collected as evidence, and a gap that shows up in at least three conversations becomes a request. Nothing is asked without evidence you can open.
What kinds of request are there?
Knowledge (something to know), Data connection (something to look up in one of your systems), Permission (something it may or may not do, such as offering a coupon), Behaviour (a rule for how it talks), and Fix (an action that failed, such as an intent that kept running without a required parameter).
What happens when I answer?
The answer becomes the right kind of thing on the agent: a knowledge entry, a row in a table, an attached document, an intent with a connection to your endpoint, a setting, or a guideline. The agent uses it from the next conversation, and the request moves to Done with a link to the version it created.
Can it apply answers on its own?
Yes, if you switch automation to Apply automatically. Drafted answers for knowledge, read-only lookups and guidelines are applied without approval, and each one creates a version with the request as the reason. Permissions always wait for a person. Off means conversations are not read at all, and Prepare answers means it drafts and you approve with one click.
Does it change my prompt?
No. Behaviour answers are saved as guidelines, which the platform folds into the generated prompt. Your prompt text is never edited by the agent. If the agent runs on a custom prompt, the page tells you that guidelines are not read until the rule is added to the custom prompt.
What is a Fix request?
A request created without any language model, from things that measurably failed: an intent that ran six times without a parameter it needed, a booking that failed, a call transfer that failed, or a procedure step pointing at an intent that no longer exists. The answers are concrete: make the parameter required, add a guideline, point the step at another intent, remove the step, change the transfer number, or mark it as handled.
How do I know an answer helped?
Each answered request shows how many conversations have used it since the date you answered, and the page keeps a monthly count of what the agent learned and how many conversations it helped.
Is customer data sent anywhere for this?
Personal details are masked before a conversation is analysed. The analysis itself runs on the same regional infrastructure as the rest of your workspace.
Let one agent ask
Open an agent, open Auto-Improvement, and switch automation to Prepare answers. Come back tomorrow to a short list of what it needs, each item with the conversations that caused it. Available on every plan.
Register free →