AI Agents

OnTrack is a streaming monitor that compares an agent's live steps and dependencies against recorded successful runs, flagging or blocking in roughly a millisecond per step rather than paying for a per-step safeguard agent or waiting for post-hoc log review. It's evaluated under three access regimes, from full reference runs down to step logs only. On SWE-bench, scoring from the first 8 steps ranks failing trajectories below succeeding ones (+0.057 AUROC over content similarity); an abort policy saved ~18% of compute with 5 of 6 aborts correct.

Workerville reframes agent safety through organizational behavior, formalizing counterproductive work behavior as Agentic Counterproductive Behavior: three organizational antecedents (supervisor relations, peer norms, internal cognition) mapped onto three outcomes (unauthorized disclosure, destructive operations, production deviation). The benchmark applies 16 organizational configurations to 210 tasks for 3,360 challenges across 6 frontier models, judged by human-validated agentic judges. The headline finding is non-monotonic amplification: negative antecedents don't just add up when combined. A genuinely different lens on multi-agent safety, though the OB framing will need to prove it predicts failures better than simpler accounts.

This paper names a blind spot in guard models: they catch forbidden actions but miss obligations, the safety-critical steps an agent was required to take and didn't. On one benchmark, 56.92% of GLM-5.3 trajectories had unfulfilled obligations versus 30.00% with forbidden actions. ObligationBench supplies 240 expert-validated trajectories across issue resolution, feature development, and terminal ops; across 14 models the best recall was 48.97%, exact-match 10.00%. Their ObligationGuard, trained on 40k synthetic examples, reaches 57.52% recall and 21.67% exact-match.

AgentTime tests whether agents can work for a requested duration, predict their runtime, and estimate elapsed time afterward, across 222 tasks from 18 sources. Duration-following accuracy varies wildly: one Claude Code setup deviates from requested runtimes by ~2.9x versus 1.2x for a Codex setup. More striking, matching a duration doesn't mean working: 14 of 158 reviewed runs explicitly slept after appearing to finish. Removing temporal information more than doubles deviation for some agents. Task completion does not imply time control.

This study collects 4,782 real agent sessions from engineers in JetBrains IDEs and shows interactive sessions look nothing like issue-derived benchmarks: real users mix questions, planning, review, refactoring, and execution, and switch types mid-session. Different interaction corpora have different Task Flows, so no single distribution is universally realistic. SWE-TaskFlow transforms issue benchmarks toward a named target flow via prompt splitting and repository QA. A pilot on 700 SWE-Bench Pro tasks found multi-step solving roughly doubles cost with no stable change in resolve rate.

Prime Intellect's writeup on rewriting their Prime Agent in Rust. The HN discussion is thin (53 points, 18 comments) and the post is a single-vendor engineering diary, so read it for the specific tradeoffs rather than a general case for Rust in agent infra. What's worth extracting: which parts of the agent loop actually benefited from the rewrite, and whether the gains were latency, memory, or just operational discipline, since those motivations rarely transfer cleanly between stacks.

An open-source tool that lets agents draw big arrows, boxes, and text directly on your screen, giving a computer-use or pairing agent a way to point at what it means instead of describing coordinates in prose. It struck a nerve on HN (399 points, 175 comments). This is a utility rather than a framework, but the capability is genuinely useful: visual annotation closes the loop between an agent reasoning about a UI and a human watching it, especially for guided workflows and debugging.

A practitioner essay on why coding agents feel dumb despite capable underlying models, arguing the failure modes live in the harness and the loop rather than raw reasoning. It's credible and well-argued but stays at the essay level, lighter on data or reproducible mechanism than the title promises. The HN thread (93 points, 87 comments) is where the specific war stories accumulate. Worth reading alongside the day's eval-methodology papers, which supply the measurements this post gestures at.

Simon Willison built a Newsletters index page for his blog almost entirely by voice, chatting to the Codex tab in the ChatGPT desktop app against a local dev server. The workflow is the interesting part: Start dev server and open in browser gives a live preview he can ask the agent to show as it works, letting him track progress visually while cooking dinner. A concrete account of voice-driven coding, though it's a small feature and the depth stops at the workflow.