> ## Documentation Index
> Fetch the complete documentation index at: https://help.ciarem.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# The playground: test and train your agent
> Simulate real conversations, read the decision analysis behind every reply, and turn corrections into lasting behavior.
**Test agent** (also reachable as **AI training** from the Control center) opens the playground, a full simulation of a real conversation:
* Pick the **channel** to simulate (a WhatsApp conversation formats differently than web chat), switch to a **simulated time** to test out-of-hours behavior, **impersonate an existing contact** to test with their real data, and choose which of your **agents take part**.
* Chat as if you were the customer, or use the quick scenario chips (upset customer, pricing question, wants to book, hours question).
* Every reply comes with a **decision analysis**: the intent it detected and the routing confidence, which agent stepped in and why, which knowledge bases and catalogs it consulted, and which transfer rules were evaluated, including why none fired.
* Under each reply, **Good response** / **Needs correction** are the training loop. A correction lets you either **improve the response** in plain language or say it **should transfer**, and the agent learns it for every similar situation going forward. The same corrections flow powers the [Evaluation center](/ai-agent/evaluation-center) for replies already sent to real customers.
## Reading the decision analysis
| Line | What it tells you | If it looks wrong |
| ------------------------------------------------- | -------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Assigned agent** and **Routing confidence** | Which agent took the message and how sure the router was | Low confidence or the wrong agent: the agents' *when should this agent step in* texts overlap, or don't match how customers write. See [Agent team](/ai-agent/agent-team) |
| **Knowledge used** / **Answered without sources** | Which sources the reply drew on | A business question answered without sources is a gap: the fact is missing from the [knowledge base](/ai-agent/knowledge-bases), or that base is not switched on for the agent |
| **Transfer rule** | Which rule fired, if any | Fired too early, or not at all: reword the condition. See [transfer rules](/ai-agent/human-handoff-and-escalation) |
## What a correction can and cannot teach
**Needs correction** opens two options, and they teach different things:
| You write… | What Ciarem does with it |
| --------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Improve response**: "Don't confirm coverage you haven't checked. Ask which plan they have first." | Folds it into the agent's standing coaching: a short list of behavioral corrections, merged and generalized to similar situations. |
| **Improve response**: "The cleaning costs \$50, not \$40." | Detects a business fact and refuses to store it as behavior. You see **Business information detected** with a link to the knowledge base. Fix the fact there. |
| **Improve response**: "Be warmer" / "Use a formal register" | Tone and personality are not coaching either: set them in [Basic configuration](/ai-agent/basic-configuration). |
| **Should transfer**: a name and "the customer disputes a charge" | Creates a real [transfer rule](/ai-agent/human-handoff-and-escalation). The coaching never decides when to hand off. |
| A correction on a reply that was a rule's transfer message | Won't change it. Edit the message on the rule itself; the playground tells you which one. |
Corrections teach **behavior**. Facts go to the knowledge base, handoffs to transfer rules, tone to Basic configuration. A correction that tries to carry a price or a policy is caught and redirected. It never silently becomes agent behavior.
Every verdict also leaves a test behind: **Good response** protects that behavior, and a correction re-runs the flagged message to check that the fix holds. The [Control center](/ai-agent/agent-health)'s **AI training** card tracks how many replies you have rated.
## A routine that works
In your customers' words, misspellings included ("cuanto vale la limpieza?", "do u do weekends").
Do this before any change to agents, rules, or knowledge. Mark **Good response** on the ones that are right. Each becomes a test that protects that behavior.
Compare with the decision analysis: the same answers where you expected no change, better ones where you did.
A wrong fact: the knowledge base. A wrong behavior: **Needs correction**, then **Improve response**. A moment that should have gone to a person: **Should transfer**.
Once you are live, the [Evaluation center](/ai-agent/evaluation-center) does the same for the replies your agent actually sent.