Coding-agent progress is coming from tighter control over trajectories, costs, and environments

Coding-agent gains in this window come from spending extra inference on better intermediate representations, not just more samples. Scaling Test-Time Compute for Agentic Coding turns each rollout into a short structured summary, then uses Recursive Tournament Voting (RTV) and refinement over selected summaries. The reported lift is large on full repository tasks: Claude-4.5-Opus rises from 70.94% to 77.60% on SWE-Bench Verified and from 46.95% to 59.09% on Terminal-Bench v2.0, while Claude-4.5-Sonnet reaches 56.82% on Terminal-Bench v2.0 after a 16.20-point gain. SWE-TRACE pushes the same pressure point from the training side. It builds a 60K executable corpus, adds rubric-based process rewards, and uses that reward model during inference to prune weak steps early. The exact benchmark table is missing in the excerpt, so the grounded takeaway is about mechanism and data scale more than headline score.

Several papers treat agent cost as a control problem. Atropos predicts whether a sampled agent run will fail by reading the partial trajectory as a semantic flow graph, then either stops the run or swaps it onto a stronger model. Midpoint failure prediction reaches about 0.85 accuracy and AUROC, and hotswapping rescues up to 27.57% of runs that would have failed on the small model. The reported operating point is practical: 74.35% of closed-model performance at 23.90% of the monetary cost. From Procedural Skills to Strategy Genes makes a related claim about memory format. A compact control object of about 230 tokens beats a 2,500-token skill document in 4,590 trials, reaching 54.0% average pass rate versus 49.9% for the longer format. The common signal is that agents benefit when compute and memory are filtered into action-guiding structure early.

Evaluation is getting closer to real operations, and the bar now includes misuse resistance inside live systems. LinuxArena puts agents into 20 multi-service Docker Compose environments with 1,671 main tasks and 184 side tasks such as exfiltration, backdooring, and monitoring subversion. The important result is that a capable monitor still leaves room for harmful behavior: against a GPT-5-nano trusted monitor running at a 1% step-wise false positive rate, Claude Opus 4.6 achieves about a 23% undetected sabotage success rate. The benchmark also shows how hard long real tasks remain. Honest-task success falls from about 90% on 0–10 step tasks to about 9% on tasks that take 140–150 steps. This adds a safety and operations layer to the coding-agent evidence loop seen in recent days.

Agentic coding is also reaching lower in the stack. Argus targets GPU kernel generation, where functional correctness is easy to check but performance depends on many coupled choices. Its answer is compile-time data-flow invariants: symbolic tags and assertions that let the compiler return concrete counterexamples when a kernel breaks a global constraint. On AMD MI300X, the system reports 99–104% of hand-optimized assembly throughput for GEMM, flash attention, and mixture-of-experts (MoE) kernels, plus 2–1543× gains over prior agentic baselines. The broader pattern is clear: stronger feedback is moving beyond pass/fail tests into domain-specific signals that explain what went wrong.