RESEARCH

The text agents read is a control surface: bad MCP error messages cost more on stronger models, and context hygiene gains value with every model upgrade.

Approve effects, not invocations: why agent approval records miss harm in install hooks and skill combinations, and why pipelines need closure metrics.

CodeMidas turns existing repository code into executable coding-agent RL tasks, tests, and verifiers, instead of relying on issues, commits, or hand labels.

An agent's report on its own run is unreliable, and eight studies show why. The fix: append-only traces, obligation ledgers, and records the agent cannot write.

AI scribe audits show omission is the dominant error, and LLM judges are near coin-flip at catching it. The fix: enumerate the facts, then check each one.

Coding agents need better human loops: a position paper argues agent research over-optimizes solo autonomy while developers need to steer and verify agents.

Agent instruction following collapses past five or six simultaneous constraints. Why rules also die in compaction and handoffs, and why side channels help.

AI agents often claim completion they did not earn. New benchmarks measure the gap, and evidence-carrying termination makes done a checked artifact.

Agent model compression has a hidden tax: quantized and pruned agents lose knowledge first and stay confidently wrong. Why cheaper agents start at the cache.

Self-improving test generation loops can grade against an oracle that agrees with itself. An audit shows the inflated gain and why independent checks count.

LLM judges flip verdicts under pushback and rankings shift with token budget. The design spec for a judge system, including cheap deliberating juries.

DeepSeek Harness (dsh) code read: a 453K-line plugin runtime on a vendored Cordis fork, three patterns worth stealing, and the gaps the launch does not mention.

A new August 2026 paper argues that coding agents improve faster when they compare attempts across tasks and lineages, not just retry one failed trajectory.

Skill files are a supply chain attack surface: adversarial skills exploit a mainstream coding agent in over 95% of runs. The fix is admission engineering.

ACE vs ALTK-Evolve for agent memory: both turn trajectories into lessons, but calibrated delivery matches accuracy on AppWorld at a fraction of the tokens.

Encrypted chain-of-thought from Anthropic, OpenAI and Google APIs can be replayed into weaker models and read verbatim. How the attack works and what leaked.

A new arXiv paper probes six widely used open-source agent frameworks and finds the barrier semantics of approval gates, cancellation, and timeouts hold on none of them. A sibling branch can execute while the user is rejecting another one, and replay can double-execute. The fix is a verified external gate called SoundGate.

AgentChaos injects server errors, truncated responses and corrupted tool calls into live agents. Pass@1 drops up to 50 points: robustness is a systems problem.

LivePlan monitors SWE-agent runs and calls an advisor LLM only on drift, raising resolution rates up to 15.2 points for an extra $0.08 per instance.

WeatherNext Cyclones, DeepMind's open-source tropical cyclone model, adds a day of forecast lead time. What the Nature paper shows and how to run the weights.