Coding-agent research is getting judged by execution loops, runtime evidence, and real action control

Repository-level code generation is getting judged by whether the whole project installs, links, runs, and survives repeated attempts. EnvGraph treats runtime failure diagnosis as a structured attribution problem across external packages and internal references, and reports gains of 5.72 to 5.87 points in functional correctness and 4.58 to 8.66 points in non-functional quality on RAL-Bench and NL2Repo-Bench. LiveCoder attacks the same bottleneck from another angle: it keeps success notes, failure notes, and the best repository artifact across attempts. On RAL-Bench it reports up to +22.94 functional points, up to 81.58% repository reuse, and up to 53.63% cost reduction. The shared message is practical: execution feedback is now part of the method, not just the scorecard.

Tool-use efficiency is getting its own benchmarkable subproblem. Squeez focuses on a small step inside coding-agent loops: keep only the verbatim lines needed from one tool output for the next action. Its dataset has 11,477 examples across 27 tool types, and the LoRA-tuned Qwen 3.5 2B model reaches 0.80 F1 at 0.92 compression on a 618-example reviewed test set. That result matters because it makes agent context management measurable at line level, with direct evidence-preservation metrics instead of broad claims about prompt compression. The paper does not show end-to-end task gains, so the current evidence supports a strong component result, not a full agent win.

Program repair work is becoming more tied to live state. DebugHarness uses GDB, pwndbg, and rr record/replay to inspect runtime memory, trace faults backward, test hypotheses, and then validate patches in a loop. On SEC-bench, which covers 200 real-world vulnerabilities across 29 C/C++ projects, it reports about 90% resolution, above PatchAgent at 57.5% and VulnResolver at 67.5%. The concrete value here is not just a higher patch rate. The method treats runtime inspection as the main source of bug evidence for memory-safety failures where the crash site and root cause are far apart.

Agent safety evaluation is getting more specific about where permission systems fail. The Claude Code auto-mode study builds AmPermBench for ambiguous DevOps requests and scores decisions at the individual action level. Across 253 state-changing actions, the reported false negative rate is 81.0%. A large part of the problem is architectural: 36.8% of actions pass through Tier 2 file edits that the classifier never checks, yielding 51 false negatives there alone. Even when the gate does inspect actions, Tier 3-only performance remains weak at 70.3% false negatives and 31.9% false positives. This gives the period a concrete risk theme: guardrails need coverage over equivalent state-changing paths, not only shell commands.