Coding-agent progress is coming from tighter feedback loops and harder evidence

Agent improvement work is getting more specific about where to intervene. CollabCoder treats debugging as a choice between fixing the plan and fixing the code, then uses stored failure history to avoid repeating weak repairs. On Qwen2.5-Coder-32B it reports 82.50 average Pass@1, ahead of CodeSIM at 80.22, with fewer API calls. A separate compiler study reaches a similar conclusion at a lower level: better feedback channels matter. On TSVC, adding compiler remarks lifts Intel success at temperature 0.8 from 2.38% to 6.95%, and hand-written dependence remarks add much larger gains. The common signal is that coding agents improve when the loop carries explicit diagnosis, not just another retry.

Repository context work is no longer only about fitting more tokens. The strongest result here says compressed context can beat full-context inference when the compression filters noise well. In the repository compression study, text-to-vector methods at 4x compression raise Python completion BLEU from 32.21 to 41.34 on QC-7B, while also cutting latency. That fits with a practical product signal from the same day: systems are trying to carry richer project context into coding tools, but the useful unit is structured, task-relevant context rather than raw prompt length.

Memory is being treated as reusable operating knowledge, not just stored traces. Memory Transfer Learning retrieves prior experiences from other coding benchmarks and finds that abstract 'Insight' memories work best. On GPT-5-mini, average Pass@3 rises from 0.523 to 0.560 across six benchmarks, with larger gains on ReplicationBench and MLGym-Bench. The paper also says algorithmic strategy transfer explains only 5.5% of gains, which points to a narrower but useful kind of reuse: validation habits, safe edit patterns, and environment-aware debugging steps.

Evaluation papers are pressing harder on generalization. One study estimates memorization advantage by perturbing inputs and tracking performance drops across 19 benchmarks. It finds test generation among the most sensitive settings, around 0.4 to 0.7, while code summarization stays below 0.3. Another paper makes the same weakness visible with real codebases: all four tested models hit 100% mutation score on open-source LevelDB whole-suite generation, yet on proprietary SAP HANA the best source-only mutation score is 10.25%, rising to 25.14% with added dependency context and still below a reduced human baseline of 30.41%. This day’s evidence favors tougher, less leak-prone checks over headline benchmark wins.