AI Agents

Under one GUI-MCP harness on OSWorld-MCP (309 tasks), the same MCP tools improve a reasoning model by +4.0pp and degrade a non-reasoning one by -5.9pp — both beyond 2 SE. The reasoning model still only calls a tool on 55/309 tasks; they name this the adoption gap. A dense tool bonus lifts spreadsheet adoption 0.03→0.33 but held-out accuracy doesn't follow: behavior is steerable, competence isn't. Dropping redundant post-tool screenshots cuts input tokens ~a third and, after retraining, beats the uncompressed operating point (37.8% vs 33.0%) at 53% of the cost.

A machine-checked conformance contract for what "resume" means when an agent workflow crashes mid-effect. Six properties (exactly-once effects, fork determinism, consume-once, and more) are checked exhaustively in TLA+ (7.4M states), then a deterministic LLM-free harness measures real frameworks at pinned releases. Findings are damning: LangGraph 1.2.9 durably records a second resume value and never reads it, persists schema-invalid state silently, and re-executes recorded work after a real SIGKILL. CrewAI and pydantic-graph fail differently; no two share a conformance profile. Consume-once holds sequentially but fires k times under concurrent delivery.

TELLER does non-intrusive root-cause analysis for LLM inference across the whole stack — inference engine, Python/C++ backend, CUDA APIs, GPU kernels, and distributed comms — by collecting NVTX/CUPTI traces and service logs without touching model binaries. It reconstructs per-request call-chain trees, aligns logs to execution steps, and compresses them via a Trace Pair Encoding tokenizer that cuts per-step trace length by 80%+. A multimodal model then localizes abnormal steps and suspicious operators with explanations. Solid systems work if you operate multi-node inference and currently stare at raw timelines.

This splits incident response into two scales: a decision-theoretic rollout planner allocates security resources tactically, and a lightweight LLM agent translates that strategy into executable commands operationally. A digital twin does the heavy lifting — simulation for tactical planning, emulation for execution — which sidesteps the reliability problem of repeatedly invoking an LLM to build a full plan. Evaluated across three attack scenarios with measured recovery improvements. The architecture is the interesting part: use the LLM where its knowledge helps, not as the planning engine itself.

Iris reframes ML-engineering agents around an evolving information state rather than a tree/graph/chain search over candidate solutions. Its inquiry-revision loop uses epistemic actions to probe decision-critical unknowns without touching the retained solution, and synthesizes observations into revisable claims with explicit scope and status. On MLE-Bench it hits a 64.9% any-medal rate under a 12-hour budget, the best among compared systems, and generalizes across four tasks spanning harness engineering and post-training. The information-vs-solution framing is worth the read even if you don't adopt the whole loop.

The Rust project is adopting an official policy on LLM-assisted contributions. The details are behind the link, but the governance signal matters: a major, safety-conscious open-source project setting explicit ground rules for how AI-generated code and reports enter its pipeline. This follows a wider pattern of maintainers pushing back on unvetted agent output. If you run code agents against upstream projects, expect more of these policies to constrain what you can submit and how you must disclose it.

A repository of agent "skills" that encode team coding standards for Claude Code and Codex, so the agent picks up your conventions rather than generic defaults. The pattern is the point: rather than repeating style rules in every prompt, you package them as reusable skills the agent loads. Useful if you're trying to make agent output pass review without hand-editing every PR. Judge it by whether the skills actually constrain behavior or just add context the model ignores.

LLM 0.32 is the biggest release since launch. Reasoning models now stream their traces to stderr (suppressible with -R), keeping piped stdout clean. It adds server-side provider tools, redesigned content-addressable SQLite logs, and features built on OpenAI's Responses API. Tool calls, tool results, and reasoning now flow as typed streaming events. If you script against LLM in pipelines, the stderr/stdout split and the log schema change are the parts to read carefully before upgrading.

The companion plugin adds Claude 5 models (fable, sonnet, opus) and server-side tools for WebSearch, WebFetch, CodeExecution, and AnthropicMCP, exposed through LLM's -T interface. This is a breaking change: the old -o web_search* options are gone in favor of -T WebSearch. Extended thinking is simplified to thinking and thinking_effort (low through max), with Claude 5 thinking by default unless you pass -o thinking 0. Narrow but concrete; the removed options will break existing invocations.

An external reconstruction of how ChatGPT Work implements Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills, and Tools. The value here is a teardown of a consumer-scale agent's architecture from the outside — how scheduling and proactivity are wired, how memory is scoped, how the tool surface is organized. Treat it as informed inference rather than documentation; it's a reconstruction, not a spec. Worth reading if you're designing comparable subsystems and want a reference point at scale.