Test before you publish

Intermediate · ~10 min · Prerequisites: admin access to AgentForce

What you'll do

Use the builder's test panel to run real conversations against your draft agent, verify that tools fire correctly, and iterate on the draft until it's ready to publish.

Open the test panel

In the builder at /admin/agents/<slug>/edit, click the Test button in the top-right corner. The test panel slides in from the right side of the screen and gives you a full chat interface connected to your current draft.

Test conversations are sandboxed — they don't appear in members' conversation histories, and they don't count toward the agent's analytics or usage metrics. You can test freely without side effects.

See Testing an agent for everything the test panel tracks.

Run real prompts

Don't test with generic input. Use the actual prompts your members will send — copy them from Slack, support tickets, or wherever the real work comes from. Run at least three prompts that cover:

• A typical task the agent is designed for
• An edge case or ambiguous request
• A request that falls outside the agent's scope (to confirm it handles out-of-scope gracefully)

Check each response against your instructions: does it follow the role statement? Does the output match the format you specified? Does it respect the rules and boundaries you listed?

Check tools fire correctly

If your agent has tools enabled, the test panel shows a tool-call trace for each response. For each tool you've added:

1. Send a prompt that should trigger that tool.
2. Confirm the tool appears in the trace and returns a result.
3. Check that the agent uses the result correctly in its response.

If a tool doesn't fire when it should, go back to the Tools section of the builder and confirm it's enabled for this agent. If the agent calls a tool when it shouldn't, tighten the instruction that describes when to use it.

Iterate on the draft

Testing almost always reveals something to improve. The iteration loop is fast:

1. Spot an issue in the test panel — wrong format, missed rule, tool misfiring.
2. Click into the Instructions or Tools section of the builder and fix it.
3. Click Save — the test panel picks up the updated draft immediately (no page reload needed).
4. Re-run the prompt that exposed the issue.

Repeat until the agent handles your three test cases cleanly. You don't need to achieve perfection — aim for a stable baseline. You can improve the agent after publishing without members ever seeing a broken version.

Try this next