RESEARCH

A Microsoft Research analysis of 3.2M users and 761M LLM calls shows coding agent traffic is 87% agent-initiated, burns KV cache at turn boundaries, and punishes every tool failure with up to 4x compute.

Evidence gates, verifiable reward games, deploy-time certificates: the fixes that moved agent quality this week did not make judges better, they removed the judge. We think the LLM verdict inside the agent loop is a transitional technology, and here is the bet you can grade us on.

Since we published your-benchmark-is-lying-to-you, roughly 25 new results have landed on the eval-integrity question. The surprise: every fix that works is structural - ledgers, counterfactuals, decompositions, personas, readout discipline - and none of them asks the model to be smarter.

A wave of audits in the last two days measured the noise floor of agent benchmarks: misaligned ground truth, lenient model judges, and aggregate scalars that hide real failures. Here is what the numbers actually mean, what to trust, and how to buy agents without being played.

A late-July research wave - native in-backbone memory, pretrained parametric memory at scale, memory reconstruction, and transactional memory writes - challenges the external-store paradigm every agent memory product is built on. Here is what changes by late 2027 and what developers should do now.

A placebo-controlled study on MBPP+ finds that when small code models fail, resampling from scratch beats repair loops that feed the failed code back - at 2.5-5.5x fewer tokens. The failed attempt is the anchor.

A new 600-session benchmark shows coding assistants that read a user's resolved session history resolve ambiguous requests with far fewer clarifying questions - Claude Opus 4.8's first-turn success jumps from 24.3% to 60.3% when history is available.

Microsoft's Change2Task turns merged pull requests into verified, executable coding agent tasks: 79.6% construction success across 1,130 repo changes, 29.2% more verified tasks than PR baselines, and tasks that stay current with the codebase.

DeepSeek re-post-trained V4 Flash into an agent workhorse: Terminal Bench 82.7, DeepSWE 54.4, native Responses API, and first-party Codex support - all at $0.14/$0.28 per million tokens. What changed, what the numbers actually mean, and how to wire it up today.

A July 2026 Microsoft study of Claude Code and GitHub Copilot CLI found roughly 24% more merged pull requests among adopters, but the interesting lesson is rollout design, not magic productivity.

A new benchmark drops five frontier coding agents into a live OpenTelemetry microservice system with real Prometheus, Jaeger, and OpenSearch telemetry. Best RCA accuracy: 25.3% on Medium, 10.0% on Hard. Even Claude Fable 5 is far from oncall-ready.

A new paper wraps code into an OWL2 ontology with SPARQL property paths to answer multi-hop structural queries for coding agents - 2.06x retrieval recall and 28.8% fewer tokens on SWE-bench Pro, versus treating code as plain text.

A systematic audit of SWE-bench Verified finds 68 of 500 instances (13.6%) pair a pull request with an issue it does not actually resolve, penalizing agents that correctly solve the stated problem. PAIChecker, a three-phase multi-agent checker, flags them with up to 92.12% binary accuracy.

A Michigan team measures prose SKILL.md files against compiled harnesses: agents execute only 56% of the steps their own skill mandates. SIGIL compiles skills into typed graph harnesses, hitting 86% compliance with 0.58x the tokens.

A new 188-task benchmark for non-functional improvements finds coding agents hit 70% on functional correctness but lag humans on refactors and structural changes - the quality gap that becomes tech debt.

A thought experiment with the sci-fi removed: freeze the models at today's capability, drop the price to zero overnight, and work out what actually changes for a working developer. Less than you fear, more than you think, and not where you expect.

FermiSense fine-tuned Qwen 3.5 9B with 2,500 GRPO steps on a single GPU for $500 and beat GPT-5.6 Sol (93%) and Opus 4.8 (91%) on automotive catalog review, reaching 97% accuracy at 68x lower cost per listing.

Black Forest Labs released FLUX 3, a single multimodal model trained jointly on images, video, and audio that also drives robots on Audi production lines. Here is what it does, how it works, and how to try it.

Terence Tao published a deep mathematical digestion of the Jacobian conjecture counterexample discovered by Claude Fable 5. Here is what happened, what HN is saying, and what it means for AI-assisted research.

ByteDance's Dockerless paper asks whether coding-agent patches can be verified without spinning up per-repo environments. The practical answer is not replace CI. It is use cheaper evidence before CI.