𝗗𝗼𝗻𝘁 𝗨𝘀𝗲 𝗗𝘂𝗺𝗯 𝗠𝗼𝗰𝗸𝘀 𝗙𝗼𝗿 𝗟𝗟𝗠 𝗔𝗴𝗲𝗻𝘁𝘀

𝗗𝗼𝗻𝘁 𝗨𝘀𝗲 𝗗𝘂𝗺𝗯 𝗠𝗼𝗰𝗸𝘀 𝗙𝗼𝗿 𝗟𝗟𝗠 𝗔𝗴𝗲𝗻𝘁𝘀

You want to load test your LLM agent. You create a fake API. It returns one hardcoded response. Your agent loops forever.

Agents need context. They call a tool. They see the result. Then they give a final answer. A dumb mock ignores the result. It sends the same tool call again.

Turn-counting fails too. Some paths use one tool. Some use three. Hardcoded turns break when the path changes.

The fix is simple. Look at the message history. Is the last message a tool result? Return text. Otherwise, return a tool call.

This approach reveals the truth:

• Infrastructure stays healthy.
• Memory stays flat.
• Event loops stay fast.
• The bottleneck is usually the API rate limit.

Dumb mocks hide this. They make your system look faster than reality. They hide the real ceiling. Build a mock understanding the protocol. Isolate your hosting from model latency.

Source: https://dev.to/sravan_vidiyala/you-cant-load-test-an-llm-agent-with-a-dumb-mock-2o7e Optional learning community: https://t.me/GyaanSetuAi