Coding-agent progress came from the interface around the model

Repository agents gained most when file reading, patch writing, and repair were split into smaller jobs. SWE-Edit keeps exploratory file content out of the main reasoning context by using a Viewer for relevant code blocks and an Editor for patch execution. On SWE-bench Verified, it reports a resolved-rate gain from 69.9% to 72.0%, edit success from 93.4% to 96.9%, and total inference cost down 17.9%.

SAFEdit applies the same division to instructed edits. A Planner writes the edit plan, an Editor changes only the target code, and a Verifier runs real unit tests with up to three repair rounds. On EditBench, it reports 68.6% task success rate, ahead of its GPT-4.1 ReAct baseline at 60.0%. A Claude Code regression report shows the operational risk on the other side: a repeated malware reminder inside Read and Grep results caused 3 of 5 Opus 4.7 subagents to refuse ordinary refactor tasks in one reported workflow.

The harness around a coding model is now treated as an editable system with its own measurements. Agentic Harness Engineering (AHE) exposes prompts, tools, middleware, skills, sub-agent settings, and memory as files. It uses cleaned rollout traces to propose changes, records predicted gains and risks, then checks outcomes before keeping edits. After 10 iterations, AHE reports Terminal-Bench 2 pass@1 rising from 69.7% to 77.0% across 89 tasks. Its frozen harness also uses fewer tokens on SWE-bench-verified than the seed harness.

Mesa and the agent security post show the same concern in product form, though with weaker measurement. Mesa gives agents a durable POSIX-compatible filesystem with branches, diffs, rollback, audit trails, and scoped mounts. The security post keeps real credentials outside an AI SRE container by injecting them through HTTP proxies and considers gVisor network interception when applications ignore proxy settings.

IssueSpecter extends automated software repair earlier in the pipeline by asking a large language model (LLM) to inspect code that current tests never execute. SlipCover finds uncovered Python segments. GPT-5-mini generates candidate issue reports with severity, reproduction steps, affected operating systems, and suggested fixes. A ranking step then filters issues for human review.

The reported scale is large enough to be useful but still noisy. Across 13 active Python projects, the system produced 10,467 issue reports. Human review of 130 top-ranked reports found 49 valid bugs, 61 that needed further investigation, and 20 invalid reports. The authors also report that LLM-based ranking beat rule-based ranking by 50% at precision@3 and 41% in mean reciprocal rank.

Efficiency claims were concrete in software-engineering models. Carbon-Taxed Transformers (CTT) combines neural architecture search, structured pruning, quantization, and distillation for code-focused models. It reports up to 49× memory reduction, 8–10× lower latency for clone detection, 4–7× lower latency for generation, and up to 81% lower inference CO2 emissions. Code generation is the hardest case: pass@1 retention reaches up to 68%.

A separate observability survey argues that production LLM systems need connected signals across model internals, confidence, behavior, operations, and infrastructure traces. Its evidence is uneven because it summarizes several studies, but the operational point matches the agent papers: failures and costs are visible only when model behavior is tied to tools, traces, and infrastructure metrics.