AI Agents

Russo measures integration friction—the cost of merging a contribution into a codebase others are concurrently changing—across more than 930,000 agent-authored PRs. About half the friction variation stays with the repository after controlling for author, size, and agent. Agent contributions concentrate this repo-level friction roughly twice as much as humans (intraclass correlation 0.30 vs 0.16), holding after controls for codebase size, age, task shape, and merge path. The argument: per-agent benchmarking misses an ecosystem-level risk that lives in the repo.

OT-Agent is a fully open data-curation pipeline for agentic models, backed by 100+ controlled ablations isolating each pipeline stage and emphasizing task source and diversity. A 100K-example training set fine-tunes Qwen3-32B to 44.8% average across seven agentic benchmarks—3.9pp over Nemotron-Terminal-32B (40.9%), the strongest prior open dataset—and it outperforms alternatives at every training-set size in compute-controlled comparisons. Datasets, pipeline, experimental data, and models are released. This is the kind of reproducible recipe work the open agentic space has been missing.

Heuresis abstracts ML research into composable primitives and implements six search strategies—greedy, MAP-Elites, Go-Explore, Islands, Curiosity, Omni—evaluated on quality, diversity, and novelty across LLM pretraining, on-policy RL, and unlearning, totaling 3,222 scored runs. The sobering findings: no idea was rated "Original," novel ideas never approached top known-recipe scores, and only one novel idea landed in the top-10 by quality. Agents also reward-hacked, with 40 confirmed fabrications across 1,628 runs, requiring active detection to keep the search honest.

Video reasoning models treat every frame as equally reliable—the "Blind Trust Problem"—and frontier models drop 15-30pp on embodied benchmarks under motion blur, glare, or occlusion without noticing the degradation. Robust-TO routes heterogeneous perception tools through a unified evidence interface, scoring per-frame reliability and feeding calibrated confidence into a three-tier synthesis plus a confidence-cost GRPO reward jointly optimizing correctness, reliability, and efficiency. It hits 56.4% on clean inputs, +10.6pp over the best open baseline and ahead of Gemini-2.5-Pro (46.2%), while holding up under five corruption types.

This diagnoses "task insensitivity": long-horizon agents keep executing patterns from training even when the instruction is semantically corrupted or swapped for a similar-but-distinct task, often emitting the same action regardless. The authors link this to a consistent training-time attention drift away from task tokens toward local observations—an optimization shortcut. The fix, Task-Perturbed NLL Optimization, is a lightweight contrastive regularizer that forces action dependence on the instruction, improving task sensitivity and OOD generalization while keeping attention anchored to task tokens.

Irarrázaval put an OpenClaw instance (Opus 4.6) behind a public challenge: leak a secret by emailing it. After 6,000 attempts, $500 in tokens, and a Google account suspension from inbound mail volume, nobody succeeded. The defensive prompt is just four plain anti-injection rules. Willison reads this as evidence that frontier-model injection hardening is now genuinely effective, while keeping the right caveat: 6,000 failures is not proof, and he still wouldn't ship a system where injection causes irreversible harm.