Multiagent AI uses multiple LLM instances to check, critique, or refine one another’s outputs. Here it appears as a mitigation strategy for hallucination, especially in software tasks. One model can generate code while another tests it, then feedback loops back to the coder for revision. section 2.13
A simple workflow looks like this:
- The coder LLM generates an answer or code. section 2.13
- The tester LLM evaluates it and points out problems. section 2.13
- The coder revises the output using that feedback, and the cycle repeats. section 2.13
The appeal is straightforward: a second model can catch mistakes the first model misses. The catch is equally straightforward: the checker can fail too, or invent bad tests, so the setup is helpful but not magic. section 2.13
