Repository-grade coding exposes where agents still break

Backend generation, enterprise migration, and architectural repair all show the same weak point: agents can produce runnable code that fails engineering constraints.

Constraint Decay fixes one OpenAPI contract across 80 backend tasks and varies architecture, database, and object-relational mapping requirements. Capable configurations lose about 30 percentage points in assertion pass rate under full constraints. Database choices cause the largest losses.

ScarfBench makes agents migrate Java applications among Spring, Jakarta EE, and Quarkus. Only 1 of 204 directed migrations reaches full behavioral equivalence. Even the strongest whole-application run reports 87% compile success, 40% deploy success, and 12% test success.

SmellBench adds a design-maintenance angle. Expert review finds 63.1% of hard static smell detections are false positives. The best agent resolves 47.7% of cases, and the most aggressive repair setting introduces 140 new smells.

Two benchmarks focus on information that agents cannot get from a single failing test.

TEBench asks agents to update tests after a production-code commit. Across seven configurations, affected-test identification F1 stays between 45.7% and 49.4%. Stale tests are hardest, with average F1 around 36%, because passing tests can still need semantic updates.

ProCodeBench measures proactive intent prediction from real VS Code traces and repository context. The dataset contains about 4.63 million integrated development environment (IDE) events from 1,246 developers and 5,492 annotated intent samples. The paper reports that LLM, retrieval, and agent baselines perform much worse on real traces than on simulated traces, and that repository context helps prediction.

Several papers add explicit control over how agents spend steps, choose tasks, or recover from errors.

TACT diagnoses overthinking and overacting at each agent step, then steers hidden activations at test time. It reports AUC around 0.9 for separating drift states, +5.8 percentage points average resolve rate on Qwen3.5-27B, +4.8 points on Gemma-4-26B-A4B-it, and up to 26 fewer steps to resolve.

MAS-Algorithm uses five roles for competitive programming: algorithm selection, retrieval, planning, coding, and judging. Across five Qwen models it raises average accepted-solution rate by 6.48 points. In the same study, LoRA fine-tuning on accepted solutions adds only 0.89 points.

ASTOR applies reinforcement learning (RL) across code I/O prediction, code generation, unit test generation, and commit message generation. One shared model beats the best task-specific specialist by 9.0% on Qwen2.5-Coder-7B and 9.5% on Qwen3-8B.

The day also includes evidence about code after the initial generation step.

Build-and-Find treats an agent-written repository as context for later agents. A builder creates a codebase from a hidden specification. A finder then answers questions about intended behavior and design choices using only the artifact. Compile-pass artifact-conditioned recovery reaches 98.9%, but the protocol gates inspection-effort claims on reliable recovery, so fast reading does not count when answers are unstable.

A maintenance study over the AIDev dataset tracks 508 agent-created files and 508 matched human-created files for at least six months. Agent-created files receive fewer maintenance commits and smaller relative edits. Humans still perform 83.21% of follow-up commits to agent-created files, so post-merge ownership remains largely human.