Testing an Agent
Testing an agent is more than sending a question and seeing if it responds. The goal is to discover whether it behaves well in real situations before opening the agent to end users.
How to open the test
Section titled “How to open the test”In the agent editor, click the Test Agent button in the top right. SquadOS opens the internal Hub with the agent already selected and a new conversation ready.
You can also use the Chat button in the agents list (/admin/agents) — it goes straight to the Hub with that agent.
What to test
Section titled “What to test”Cover at least these scenarios:
- simple question within scope;
- complex question within scope;
- question outside scope (the agent should refuse gracefully);
- message with incomplete information (the agent should ask before assuming);
- request that requires a tool (the AI should call the right tool);
- request that should not trigger a tool (the AI should answer from its own knowledge);
- inappropriate tone or frustrated user (the agent should keep the configured tone).
Validating tools
Section titled “Validating tools”When the agent uses a tool during the test, check in the conversation:
- called the right tool;
- passed correct parameters;
- asked first when data was missing;
- explained the result comprehensibly;
- handled errors without inventing answers.
Each tool call appears in chat with the tool name and parameters sent — useful for auditing behavior.
Adjusting
Section titled “Adjusting”If the test fails, adjust one thing at a time: prompt, model, tool, base, or advanced setting. Then repeat the same test case to confirm the improvement. Changing several things at once makes it hard to know what fixed it.
Logs and audit after testing
Section titled “Logs and audit after testing”To investigate test conversations in more detail — including tokens consumed, per-call latency, and full tool parameters — open the Executions tab in the side menu. There you see the agent’s per-turn history.