> For the complete documentation index, see [llms.txt](https://doc.aissist.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://doc.aissist.io/capabilities/continuous-improvement.md).

# Continuous Improvement

How an Aissist agent keeps getting better - teaching it from real conversations with Continuous Learn, refining assets and instructions, and the weekly improve loop.

*Last updated: October 25, 2026*

An AI agent is never finished. This page explains how to make it better every week: teach it from real conversations, fix the knowledge and instructions behind weak answers, and check each change before it goes live.

### Why it matters

Every conversation shows you something: a question the assets do not answer, an instruction that is too vague, a workflow that needs an action. Teams that review and fix a few of these each week see resolution and CSAT keep rising. Teams that set up once and walk away see them drift.

### The improve loop

1. **Find what to fix.** The [weekly report](https://doc.aissist.io/capabilities/weekly-reports.md) lists prioritized items under **To improve**, each saying what is short, what to change and why. The **Dashboard** shows which sub-agents escalate or hand over most; **Diagnosis** next to a sub-agent asks [Ask Engine](https://doc.aissist.io/capabilities/ask.md) to analyze its conversations. Alerts tell you when something moves.
2. **Understand one conversation.** Open it in **Deploy → Sessions** and click **Ask about this session** to see which sub-agent handled each turn, what was retrieved, which actions ran, and what was missing.
3. **Fix it at the source.** Update the asset, instruction, action or handover rule that caused it (see below).
4. **Check the fix.** Run an [evaluation](https://doc.aissist.io/capabilities/evaluation.md), or **Evaluate around this session**, and compare against the earlier run.
5. **Watch the next week.** The weekly report's **Changed this week** section lists what was changed, by whom, next to how the numbers moved.

### Continuous Learn: teach by example

Continuous Learn items are short example exchanges - a customer message and the reply you want - with an optional instruction. When a new conversation is like one of them, the AI retrieves the example and follows it. Use them for answers that are hard to express as a rule: tone, phrasing, the exact steps for a tricky case.

To add one from a real conversation, open the session in **Deploy → Sessions**, click **Add to Continuous Learn**, describe the scenario and the expected behavior in **Instruction**, and click **Add Example**. Manage all items in **Deploy → Continuous Learn**: search them, open one to edit its **Instruction** and review its **Interactions**, or delete it.

### Refine assets and instructions

Most weak answers come from the source, not the reply. Fix:

* **Missing or outdated knowledge** - add the exact details customers need to the asset: steps, policy exceptions, required fields, the right link. Then test with **Debug** in **Workspace → Assets**.
* **Too much unrelated knowledge** - link assets to the sub-agents that need them, so a returns question is not answered from a shipping page. In the session, look for retrieved content that has nothing to do with the question.
* **Vague instructions** - say exactly what to ask for, in which order, and when to hand over.
* **Missing actions** - if the AI keeps handing over to look something up, connect an [action](https://doc.aissist.io/capabilities/actions.md) that does it.

### With Ask Engine

[Ask Engine](https://doc.aissist.io/capabilities/ask.md) manages Continuous Learn items for you, and every change waits for your confirmation:

> Teach the AI from this conversation: when a customer asks for a replacement net, send the spare parts link and ask for the product size.

Ask Engine shows exactly what it will add, update or delete. Nothing is written until you agree in your next message. It can also list and search existing items, or tell you which examples shaped an answer.

> Study recent conversations, find knowledge gaps in the assets and weak spots in sub-agent instructions, and say where each fix belongs.

### Best practice

* Fix the source first. Use a Continuous Learn item when a rule or an asset cannot express what you want.
* Keep examples short and specific. One scenario per item.
* Review a handful of escalated or low-CSAT conversations every week.
* Evaluate after each meaningful change, before moving on to the next.

### Related

* [Weekly Reports and Alerts](https://doc.aissist.io/capabilities/weekly-reports.md)
* [Evaluation](https://doc.aissist.io/capabilities/evaluation.md)
* [Assets](https://doc.aissist.io/capabilities/assets.md)
* [Add Assets](https://doc.aissist.io/tutorial/add-assets.md)
