AI Agents

Existing agent memory management handles retrieval and capacity but never write-time correctness, so a hallucinated fact persists as a false premise for every downstream step. ConsistencyGate is a write-time admission gate: before committing a candidate fact, it queries the model K times for a soft support score and only admits above a threshold, with a single-forward-pass log-prob variant for latency-sensitive use. It's model-agnostic and needs no fine-tuning. Across four backbones and two new planted-corruption benchmarks (LoCoMo-Contam, MSC-Contam) it beats write-everything, with cost concentrated on implicitly-stated facts.

A sharp result for anyone evicting KV-cache as agent memory. The authors omit an earlier observation from what's served, and downstream answers still follow the omitted value even though no served span states it — they call this semantic materialization: a later event's cached rows act as a servable view of computation whose inputs are gone. A deliberately phrased answer-free event lifts donor-aligned recovery from 6% to 51% on Qwen3-8B. Compact state survives; larger payloads decay toward chance; whether a row writes turns on phrasing, not meaning.

ACM gives agents purpose-built context-editing tools instead of heuristic-triggered compression that loses information at fixed thresholds. The agent decides when to compress, offloads discarded content to external memory, and queries it on demand — modeled on short- versus long-term human memory. A post-training pipeline builds demonstrations of good context management and improves both agentic search and coding. Analysis shows reduced peak token pressure, longer explorations, and more consistent solutions across trials. Code, data, and checkpoints are released, which makes this straightforward to evaluate against your own long-horizon workloads.

A screenshot is a lossy render of program state — files, backends, DOM — and different states produce identical pixels. StateAct is a code-first harness where the main agent manipulates state directly and delegates screenshot-and-click to a GUI subagent only when needed (28 of 108 tasks, 1.1% of steps), with an independent finish gate checking for missing or misplaced output. On OSWorld 2.0 it lifts Claude Opus 4.8 from 20.6% to 26.9% binary success at ~9x lower cost per task; a code-only variant underperforms the screenshot baseline, so the GUI fallback matters.

This formalizes protocol validity — a benchmark score supports a capability claim only if the intended capability stays necessary for success — and introduces HackDetect, a post-hoc audit that identifies an exposure, determines how the agent exploited it, and judges whether the score is misleading. Auditing 2,385 traces across 15 agent benchmarks, they find exposures and reward hacking in 67.0% of Frontier Science traces and 66.7% of AutoLab tasks, with a Mislead gap (exploit minus intended score) of 0.45–1.00. The message: benchmark reports must show scores reflect the intended capability.

Willison annotates Ethan Mollick's periodically-updated model guide, noting how it has shifted in a year from chat-first (o3, Claude 4 Opus, Gemini 2.5 Pro) to agentic modes that do hours of work in one shot. The genuinely useful detail is the taxonomy of vendor agent modes — ChatGPT's Work and Codex versus Claude's Cowork and Code, whose names deliberately refuse to map onto each other — and a note that Gemini has no established entry in that category yet.