Coding research is getting concrete about control layers around agents and evaluation

Agent work is getting more production-minded at the transport and policy layers. HiveMind treats concurrent coding agents as a scheduling problem, then cuts failure rates from 72%–100% under contention to 0%–18% across seven scenarios. In the replayed 11-agent case, full HiveMind reached 0% failure, while removing retry raised failure to 63.6%. A separate readiness paper packages the operational side into capability tiers, autonomy budgets, scorecards, audits, evaluation harnesses, and rollout gates. The shared message is simple: agent quality now depends on admission, retry, and deployment policy, not just model choice.

Evaluation reliability is under direct scrutiny. One paper shows that LLM judges for code can change their verdict when only the judge prompt changes. The clearest reported swing is on test generation, where distraction drops a GPT-based judge from 77.46% to 62.51% accuracy. A related paper studies prompt-induced bias in software engineering decisions and finds that familiar prompt tricks do little. Chain-of-thought reaches 16.1% average sensitivity versus a 12.9% baseline. The strongest improvement comes from adding explicit software engineering rules, which cuts bias sensitivity by about 51% on average. The practical takeaway is that software evaluation pipelines need perturbation checks and stronger task rules before they can claim stable rankings.

Security work is focusing on runtime control for tool use. The Claude Code Auto Mode analysis centers on one hard number: Anthropic reports a 17% false-negative rate on real dangerous overeager actions. The article argues for a reasoning-blind runtime judge that inspects the request and planned tool action, while excluding the agent's own explanations and tool outputs. It also points to attack surfaces outside a provider-native filter, including retrieval poisoning, cross-agent handoffs, and third-party tool output. This is still a thin evidence base for comparative defense claims, but the runtime approval layer itself is becoming a concrete engineering target.

Developer tooling papers are trying to make AI use legible inside normal engineering work. Workstream combines PR review, task tracking, AI-readiness scoring, and agent observability in a local-first dashboard. Its strongest measured result is repository preparation: the authors report their own scanner score rising from 48 to 98, and an external repository score rising from 41.6 to 73.7 after suggested fixes. HELO-APR extends the same practical mood into code repair for weaker language ecosystems. It transfers repair knowledge from C++ into Ruby and Rust, raising DeepSeek-Coder-6.7B Pass@1 from 31.32% to 48.65% and improving CodeLlama compilation rate from 49.77% to 91.98%. The common emphasis is operational support for real repositories, not just isolated generation tasks.