Coding agents are being fenced with traceability, cost checks, and production validation

Rel(AI)Build targets a practical weak point in coding-agent use: the files that define prompts, permissions, and tool behavior often have little provenance or review history. Its corpus study found exact duplicate agent config paths in 10.1% of tracked cases after fork adjustment, and 75.5% of duplicate clone pairs crossed organization boundaries. The proposed control plane adds hashes, lockfiles, audit logs, permission tiers, pre-tool checks, and compilation into seven IDE targets.

Knowledge-Based Pull Requests (KPR) applies a similar control instinct to collaboration. External code, tests, logs, and cleaned agent traces become a reviewed knowledge package. A project-owned agent then regenerates candidate code inside the receiving repository. The pilot covers seven merged public pull requests, so the evidence is early, but the workflow names a real review problem: maintainers need intent, risk, and provenance before they accept agent-assisted changes.

CodeAnchor shows that simple static facts can make grep-first agents easier to inspect. It inserts call, import, inheritance, configuration, data-flow, I/O, and test links as comments next to code. On SWE-bench Lite, lightweight topology improved Func@5 by 2.2 percentage points and shortened navigation by 1.6 interaction rounds, with about 10% more input tokens.

The execution-cost study questions a common repair loop. Across 7,745 public SWE-bench traces, agents ran tests 8.8 times per task on average. In 3,000 controlled attempts, commercial agents gained only 1.25 percentage points in resolve rate under unrestricted execution, with no statistically significant gap. Claude Code resolved 63% without execution and 64% with unrestricted execution, while the no-execution setting saved 56% of tokens and 48% of wall-clock time.

Two repair papers warn against relying on a single aggregate score or a single scanner result. The quantization study finds that smaller or quantized LLMs can reduce memory by up to 85%, yet many settings raise inference time or energy use. Some quantized variants repair more bugs than the base model, including a DeepSeek-Coder-6.7B result that improved Defects4J plausible repairs from 43 to 82. Similar pass counts often came from different solved-problem sets, so the authors add a Jaccard-style consistency measure.

TerraProbe makes the oracle problem concrete for Terraform security repair. Gemini cleared the targeted Checkov finding in 83.3% of first-pass repairs, but full Checkov cleanliness fell to 10.4%. Among plan-compared real-world TerraDS repairs, 71.4% were deceptive fixes that passed automated checks while leaving the target vulnerability in place. The paper’s layered evaluation adds terraform validate, terraform plan, JSON plan comparison, and human labels.

AgentX and NOVA give the day’s clearest industrial deployments. AgentX runs a four-stage recommender loop: proposal generation, repository-grounded code changes, safe A/B rollout, and harness updates based on trajectories. In a three-week Kuaishou App deployment, three workers generated 374 ideas and 10 launchable rollouts. The reported online gain was 0.561% user app time, with guardrail-vetoed A/B feedback stored as reusable experiment knowledge.

NOVA focuses on architecture changes in an advertising recommender used by more than 1 billion users. It records candidate model graphs and feature settings, checks semantic validity before training, and writes failed directions back into the search process. The reported L3 Literature-to-Production effective pass rate was 60.0%, more than double the human expert loop baseline in the paper. Selected online tests improved GMV on three pCVR objectives by 1.25%, 1.70%, and 2.02%.