AI AGENTS

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 SCAM 2026 study of 100 top-starred repos catalogs six configuration smells in AGENTS.md and CLAUDE.md files: Lint Leakage in 62%, Context Bloat in 42%, Skill Leakage in 35%. Only 9 of 100 files were smell-free.

A new arXiv benchmark ran 6,560 sandboxed runs across Claude Code, Codex, OpenClaw, and Hermes with five LLMs. 68% of runs triggered unsafe signals, and 66% of all runs were unsafe yet still passed completion checks. Task completion does not prove an agent ran safely.

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.

A new 106-issue benchmark across 49 repositories finds frontier coding agents rarely retrieve AI contribution rules on their own - and never refuse to contribute in AI-banned repositories, no matter the prompt. Disclosure and verification can be fixed; bans cannot.

A controlled ablation across Claude Code and Codex, 17 real tasks, and 288 evaluated runs finds context-injection strategy does not measurably change correctness (bounded to under 10-15pp). The failures are implementation skill, not missing repository knowledge.

Google DeepMind's Gemini Robotics ER 2 is now publicly available via the Gemini API. It watches live video feeds to track task progress, orchestrates VLA models as tools, and coordinates multiple robots. The numbers: 57.4% progress classification, 91.3% moment finding at 0.96s offset.

GitHub's stacked pull requests went into public preview on July 30. Stacks turn one large change into an ordered chain of small, reviewable PRs with one-click merge, plus a gh-stack skill for coding agents.

The "Building abundant intelligence" essay carries real engineering numbers: GPT-5.6 Sol cut serving costs 20%, speculative decoding gained 15%, and two settings moved ARC-AGI-3 from 13.3% to 38.3% with six times fewer tokens.

An agent CLI plus a cron schedule turns recurring dev chores into background work: dependency bumps, doc freshness checks, morning briefs. The pattern, the guardrails, and where to run it - your own hardware or a cloud host.

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.

Shopify's July 30 Hydrogen developer preview update ships Vue bindings, bundled GraphQL TypeScript tooling, Shopify Inbox AI chat, and agent skills for four more frameworks. What the rebuilt toolkit means for storefront developers and coding agents.

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.

Model prices fell 80% in a single announcement this week. Run the trendline forward and the interesting question is not the price - it is what developers, teams, and the broader economy do when intelligence stops being the scarce input.