Skip to main content
Use a mock to set the response your agent gets from a backend or replace a write during a simulation. For example, return no appointment slots or make a booking call fail. Each mock belongs to one test and is saved with that test’s version. Other tests can use different answers for the same tool.

​
Add an answer

After ## Expected behaviors, add ## Mock tools. Use the exact tool name as the subheading and put the tool’s normal result inside answer:
answer can contain any JSON value that matches the tool’s real response shape. The mock returns the same configured result each time that named tool is called in the simulation. The mock does not check the tool’s arguments. Add an expected behavior if you need to check that the agent used the right caller, provider, date, or other arguments.

​
Test a failure

Use error with a string to make the tool fail:
Write the expected behavior for the agent’s response to that failure, such as:
Each JSON block must contain exactly one of answer or error. Do not add delay, status, or matching rules to the envelope.

​
Use a connection that supports mocks

Retell phone calls do not apply these mock answers. Use text mode or a web call when your Retell test depends on mocks. For LiveKit, the simulation helper is required even when a test has no mocks. It sends the agent’s spans, reports its tools, and applies the running test’s mock answers. Set EGMA_URL and EGMA_API_KEY in the worker environment. If the setup exchange fails, the helper raises NotReported and the agent session does not start. The monitor helper is for production conversations; it does not enable simulations.
Only the tools named in the test are mocked. Every other tool runs normally. Include each write you need to replace, and verify the supported connection before running a test against a production backend.

​
Check the result

Run egma push, then start a run. Open the simulation’s transcript and inspect the tool call, its arguments, and its result. The agent’s trace records both real and mocked tool calls. Calls answered by the test’s mocks are marked mocked. The transcript prefers the agent’s tool evidence when it is available. A misspelled mock name does not match a tool. It creates no tool call or warning, and the correctly named tool can still run its real implementation. If a real backend was called, check the exact tool name, connection type, and LiveKit setup. If the agent reacts incorrectly to the answer, compare the mock’s response shape with the tool’s actual contract.