Skip to content

README.md

Latest commit

 

History

History
11 lines (11 loc) · 675 Bytes

README.md

File metadata and controls

11 lines (11 loc) · 675 Bytes
Raw

CodingAgent

AG2 coding agent on HumanEval dataset Three AG2 agents collaborate per problem (hard 3-min wall clock):

  1. Solver (AssistantAgent) — writes a Python implementation.
  2. Tester (AssistantAgent) — writes additional unit tests.
  3. Verifier(UserProxyAgent + executor) — runs the code in a sandboxed work_dir via AG2's LocalCommandLineCodeExecutor, replies with exit status and output. Used twice per attempt: once for the agent-written tests, once for the canonical HumanEval check() test. On failure, the orchestrator retries up to 3 times, feeding the verifier's output back to the solver as feedback.