AgentDevel: Reframing Self-Evolving LLM Agents as Release Engineering
Di Zhang
TL;DR
AgentDevel reframes learning-based agent improvement as a software-release process to guarantee non-regression and auditability. It maintains a single canonical agent line, generates one release candidate per iteration, and uses flip-centered gating guided by regression signals derived from execution traces and a surface-level critic. Executable diagnosis translates symptom patterns into engineering specifications used to synthesize the RC, which is then accepted or rejected based on example-level flips (P→F vs F→P) and alignment with stated intents. Across multiple execution-heavy benchmarks, AgentDevel delivers stable gains with substantially fewer regressions compared to self-refinement or population-based search, demonstrating practical viability for deploying reliable, auditable LLM agents.
Abstract
Recent progress in large language model (LLM) agents has largely focused on embedding self-improvement mechanisms inside the agent or searching over many concurrent variants. While these approaches can raise aggregate scores, they often yield unstable and hard-to-audit improvement trajectories, making it difficult to guarantee non-regression or to reason about failures across versions. We reframe agent improvement as \textbf{release engineering}: agents are treated as shippable artifacts, and improvement is externalized into a regression-aware release pipeline. We introduce \textbf{AgentDevel}, a release engineering pipeline that iteratively runs the current agent, produces implementation-blind, symptom-level quality signals from execution traces, synthesizes a single release candidate (RC) via executable diagnosis, and promotes it under flip-centered gating. AgentDevel features three core designs: (i) an implementation-blind LLM critic that characterizes failure appearances without accessing agent internals, (ii) script-based executable diagnosis that aggregates dominant symptom patterns and produces auditable engineering specifications, and (iii) flip-centered gating that prioritizes pass to fail regressions and fail to pass fixes as first-class evidence. Unlike population-based search or in-agent self-refinement, AgentDevel maintains a single canonical version line and emphasizes non-regression as a primary objective. Experiments on execution-heavy benchmarks demonstrate that AgentDevel yields stable improvements with significantly fewer regressions while producing reproducible, auditable artifacts. Overall, AgentDevel provides a practical development discipline for building, debugging, and releasing LLM agents as software development.