Two panels build and change AI Agents by conversation instead of by filling in fields. Agent Builder creates an agent that does not exist yet, from a description of what you need. Agent Improver takes an agent that already exists and changes it from an instruction you write.
A third panel does the same kind of work from different evidence. AI Agents Optimization starts from an AI Testing run and proposes changes based on what the evaluation found, so it belongs with the testing pages rather than these.
Which one to use
| Panel | Starts from | Produces | Opens from |
|---|---|---|---|
| Agent Builder | A description of the agent you want | A new agent, or a supervisor with its sub-agents | The AI Agents page |
| Agent Improver | An instruction about an agent that exists | Changes to that agent's own configuration | The agent's card |
| AI Agents Optimization | A finished test run | Changes argued from what the run scored | A run's detail page |
None of the three writes anything on its own. Each proposes, and you approve.
Before you start
Agent Builder and Agent Improver need a language model of their own, set under Administration > Settings > Agent Builder & Improver along with the model and its reasoning effort. This is a different setting from the one AI Testing uses, so the Optimizer can be working while these two are not, and the other way round.
The same page carries Automatically archive instructions before each change is applied. While it is on, an agent's previous instructions are saved before Agent Improver replaces them, and stay readable in that agent's Instructions History. AI Testing has its own copy of the setting for the Optimizer.
Agent Improver is available in the Premium edition.
You also need the Agentic AI feature enabled for the tenant and the permission that lets you edit agents. Without them the entry points are simply absent.
Talking to the panel
All three panels share the same chat input, so what follows is true wherever you are working.
The attachment icon stages files, the microphone dictates instead of typing, and send stays disabled until the typing area has text of its own.
Attaching a file
Attach a file when the request depends on something the panel cannot see: a call flow, a policy document, a JSON payload the agent has to handle. Pick the files with the attachment icon or drop them anywhere on the panel. Each appears as a chip you can remove before sending, and a file that cannot be accepted is reported by name.
A message needs text of its own. Files on their own will not send, so say what the panel should do with them.
Attached files are not stored. They travel with the one message they are attached to and are not carried into later turns, so a follow-up that depends on the same document needs it attached again.
One thing to check first: nothing verifies that the configured model can read files. If it cannot, attaching a PDF or an image fails when the request reaches the model, and the panel reports a generic error rather than warning you beforehand.
| Item | Value |
|---|---|
| Files per message | 5 |
| Size per file | 5 MB |
| Accepted file types | PDF, PNG, JPG, JPEG, JSON, TXT, MD |
Dictating instead of typing
Click the microphone, speak, and click it again. The recognised text is written into the input at the cursor and existing text is never replaced, so a message can be part typed and part dictated. The second click does not send: read the text, correct it, then send.
Recognition runs in the language of the project's Speech Recognition integration. The microphone appears only when the project has one in use whose language matches the project's primary language, and when it is missing nothing on screen explains why, so that is the pair to check.
Next
- Building an AI Agent with AI: describing the agent you need, the questions the Builder asks, and confirming what it creates.
- Improving an AI Agent with AI: asking for a change to an existing agent, reviewing the proposals, and what happens to the agent when you approve them.
- AI Agents Optimization: the same kind of work, argued from a test run instead of from a sentence you write.