AgentDevel: Reframing Self-Evolving LLM Agents as Release Engineering

Open paper
Benchmark Model Rank Results
code-generation-on-swe-bench-verifiedAgentDevel (final) (Claude Code)–Accuracy: 30.0
computer-use-agents-on-webarenaAgentDevel (final)–Success Rate: 35.5

Rank counts only results with a code link.