AgentCoder: programmer + test designer + test executor
Multi-agent code generation with iterative testing and optimization—explicit separation of writing code and writing/running tests.
Source: AgentCoder: Multi-Agent-based Code Generation with Iterative Testing and Optimisation
AgentCoder splits roles: a programmer agent, a test-designer agent, and a test-executor loop that feeds failures back for optimization. The architectural claim is that iterative testing should be a first-class multi-agent protocol, not an afterthought tool call.
It is an early, clear example of verification-oriented multi-agent coding—closer to review/QA structure than to a single chat session that sometimes runs pytest.
Why it matters for code review
Review quality climbs when someone other than the author owns the failing tests. AgentCoder’s role split is a template: do not let the same agent that wrote the patch monopolize the definition of “done.”