Recommendations
The ranked next steps the platform suggests for an agent, what each one tells you, and how the list is kept current after every revision.
Recommendations, under Improve in the workspace sidebar, is a list of next steps for the agent, ranked by how much each is expected to help. The app describes it as opinionated: it tells you what to do next rather than listing everything you could change.
The list is re-evaluated after every revisionRevisionThe second number in a version, v0.3. Increments on every save and starts again when the agent is published.See the glossary, so it tracks the agent as you save. Opening the section evaluates it again, and so does Re-check at the top right.
Reading a recommendation
The header counts what is still open, for example 1 open. Recommendations are grouped by impact, under High impact, Medium impact and Low impact. Each one carries:
- a title naming the state the agent should reach, for example Test scenarios have been run;
- one sentence on why it matters, for example: this agent has test scenarios that were never run, so nothing has validated it yet;
- Fix it, on some recommendations, which opens Atlas with a request to fix it already written; nothing is sent until you send it;
- a dismiss button, Dismiss recommendation, which removes it from the list.
When nothing is open the section reads Nothing to recommend right now.
Work from the top. The order is the platform’s estimate of which change moves the agent most, so the first item is the one worth your next few minutes.
What it checks
Each recommendation is one of these checks failing for the agent. Which ones apply depends on the agent: a voice check does not appear on a text agent, and Flow has nodes only on a flow agent.
| Check | Why it matters |
|---|---|
| Prompt is configured | The system prompt defines how the agent behaves, so it can’t be empty. |
| Flow has nodes | A flow agent with an empty canvas has nothing to say when a call connects. |
| Opening line set | An agent that speaks first needs an opening line to start the call with. |
| Agent description set | A short description helps your team recognize this agent at a glance. |
| Agent name is unique | Sharing a name with another agent in this workspace makes them easy to confuse. |
| Voice selected | Without a voice of its own, the agent answers in the workspace default. |
| Voice matches the language | A voice that does not speak the agent’s language is heard by callers as an accent. |
| Maximum call duration set | A duration cap stops a stuck call from running, and billing, indefinitely. |
| Outcomes defined | Outcomes are what a call reports back: without them every call ends unclassified. |
| Knowledge base is usable | An empty or unavailable knowledge base returns nothing to lookups during a call. |
| At least one test scenario | Test scenarios let you validate the agent before shipping it. |
| Test scenarios have been run | Scenarios that were never run have validated nothing yet. |
How it relates to the other checks
Recommendations and the Launch checklist run the same checks; Prompt is configured and At least one test scenario appear in both. The difference is what they do with a failure. The checklist is a gate: its Required items decide whether the agent can be published. Recommendations rank everything that is still open by impact, as advice, and an agent can be live with some still open.