Evidence-rich control loops are the center of agentic coding work

Agent training work focused on the evidence available during training, not only on final task scores. EvoTrainer evolves policy branches together with the diagnostic code that reads rollouts, logs, configs, and code diffs. Its strongest reported gain is on repository-level software engineering: SWE-9B reaches 38.16 Avg@8 BC% versus 33.77 for a human-engineered reinforcement learning setup.

The terminal-agent study reaches a related conclusion through supervised fine-tuning. DeepSeek-V3.2 has a lower standalone Terminal-Bench 2.0 score than Claude Opus 4.6, yet its traces train stronger Qwen3 students because they expose more inspect-act-verify behavior. Masking observation-command links cuts the Targeted Observation Ratio from 13.4% to 5.3% and drops Qwen3-32B performance from 20.6% to 13.8%.

Code repair papers made debugging feedback more local and testable. FLARE adds line-level suspiciousness scores to execution feedback, then searches candidate edits over the top-k lines. With k=10, it improves Pass@1 across five base models on LiveCodeBench and BigCodeBench, while its diagnostic model alone reaches 67% Top-1 and 89% Top-10 localization accuracy on 100 LiveCodeBench tasks.

Neural Change Prediction uses synthetic mutations and observed output changes to learn both directions of the edit-behavior link. Fine-tuned GPT-4.1 reaches 95% accuracy on CSS change prediction, 82.6% for Python change-location prediction with single mutations, and 95% for predicting output changes from a single Python code mutation. The practical value is clear: proposed repairs can be ranked by expected behavior and then checked by tests.

SPOQ treats software work as a dependency graph. It splits an epic into atomic tasks, schedules independent tasks in parallel waves, and applies planning and code validation gates with a 95% aggregate threshold. Reported results include up to 14.3x speedup on unbounded synthetic graphs, 1.4x on a 2-slot real LLM backend, and a test pass-rate increase from 91.25% to 99.75% when dual validation is used.

A broader synthesis paper gives the organizational reason for these gates. It cites 456,535 agent-authored pull requests across 61,453 repositories, but also notes that agent-authored pull requests are merged less often and tend to make fewer structural code changes. The human role described in this corpus is specification, review, governance, and escalation during ambiguous planning or risky changes.

Developer-facing tooling targeted the operating costs around coding agents. Cross-Lingual Token Arbitrage rewrites multilingual prompts locally before cloud dispatch. On OMH-Polyglot, prompt tokens fall by 34–47% while accuracy is held or improved across gpt-3.5-turbo, gpt-4o, and gemini-2.5-flash-lite, though dollar savings vary by backend because output-token pricing can offset input savings.

Project Brain addresses repeated context setup with a small Markdown index that points to topic files. It has no benchmark, but the design is concrete: load the index eagerly, open detailed project notes only when relevant, and validate broken pointers or status drift. A separate agent-network paper adds a trust angle: capability ads should include reliability estimates, benchmark names, sample sizes, evaluation dates, version limits, expiry, tests, and reputation updates.