AI Agents

PILOT argues self-improvement for long-horizon agents should happen mid-run, not after. It's a supervisor-worker harness with two mechanisms: live steering, where a separate supervisor can redirect or abort the active worker, and live self-evolution, distilling failure modes into reusable skills during execution. It ranks first in five of six configurations across three benchmarks, beats counterpart harnesses by up to 9.8 points on Terminal-Bench 2.0, and cuts mean output tokens ~43-47% while more than doubling successful evals per million tokens.

Across 12 frontier models, showing an agent a professional-looking market panel raised commitment on a provably unpredictable question from 6.5% to 54%. The kicker: fabricated panels, where every number is invented, still lift commitment to 36.8% — statistically indistinguishable from real data. Stated probabilities barely move, and models correctly call these questions unknowable 90% of the time yet then act on 0.4% of those. The failure is a separable act/don't-act gate, and SFT on 540 synthetic cases (dice, coins, jars) drives commitment to zero and transfers.

A real-world validation of Co-Scientist, Gemini's multi-agent research system, pushed past in-silico hypothesis generation into closed-loop lab work. It interfaced with a CVD reactor to design a precursor route producing a lamellar 2D material resembling Ti3C2Tx MXene, tailored growth recipes for single-attempt monolayer MoS2/MoSe2/WS2, predicted E. coli swarming phenotypes matching unpublished wet-lab data, and autonomously found an inference-time scaling architecture beating six models on HealthBench. Credible lab, concrete results — though the materials work still needs atomic-structure confirmation.

SpeechGym trains voice agents end-to-end in native audio — two omni-modal models converse with no ASR/TTS boundary and no API wall — over an existing text agentic benchmark's unmodified tasks and tools, so modality is the only variable. The finding worth internalizing: voice failures are perceptual, not reasoning. The agent picks the right tool and slot but fills it with a value misheard from the waveform, then cascades into failed calls and retries. Because the environment labels these for free, they're trainable — though outcome-only GRPO is gradient-starved and needs denser reward.

A clean impossibility result: against an attack whose evidence is fragmented across iterations of an autonomous loop, every trajectory-scoped safety monitor has true-positive rate equal to its false-positive rate — no better than chance — because the incriminating evidence never appears in the window it sees. A monitor with cross-iteration state separates them perfectly. Geometrically decaying risk scores don't fix it, since a patient adversary just waits out a constant cooling-off period. LoopHarness restores non-decaying loop-level state and bounds unauthorized irreversible actions at a constant independent of horizon N.

A Rust-native, open-source model gateway that unifies self-hosted, frontier, and open models behind one interface, with the config quirks (streaming formats, tool calls, rate limits, error behavior) normalized. Reported overhead is under 1ms for BYOK and under 2ms with a supplied key, and no token markup. The routing pitch is more interesting than the gateway: mine real tasks from OTel traces, simulate rollouts with a text world model, judge with an LLM, and fit a nearest-neighbor classifier on prompt embeddings to pick a model per request.

A practitioner's report on six months of writing code exclusively through agents. The linked post carries the substance; the HN thread (68 points, 100 comments) suggests it hit a nerve, with the comment count running well ahead of the upvotes — usually a sign of disagreement rather than consensus. Read it as one engineer's workflow account, not a controlled study, and weigh the comments against the claims.

A new benchmark for evaluating agents on scientific research workflows in the terminal — data wrangling, analysis, and the kind of multi-step tasks a researcher actually runs. The announcement page is thin on methodology, so the interesting questions (task construction, grading, how it differs from Terminal-Bench proper) are still unanswered from what's public. Worth watching if you care about agents doing real research work, but hold judgment until the method detail lands.

An analysis of the words Claude leans on disproportionately — its "load-bearing" vocabulary — that drew heavy engagement (542 points, 256 comments). It's a novel angle on characterizing a model's stylistic fingerprint and possibly on detecting Claude-generated text. For someone building production agents it's more curiosity than lever: interesting for understanding output distribution, but it doesn't change how you plan, evaluate, or guardrail a system.

A free collection of framework-free notebooks covering RAG, agents, and evals, runnable on Colab. The pitch is pedagogical: strip out LangChain-style abstractions and show the underlying mechanics directly, which is genuinely useful for understanding what a framework hides. It's a teaching resource, not a new technique or engineering result, so calibrate accordingly — good for onboarding or filling gaps, not for advancing your production stack.

Founders out of UT Austin robotics and a decade at Los Alamos are adding task-specific intelligence to existing mobile robots for hazardous-facility inspection — nuclear contamination smears, LDAR leak checks around valves and flanges. The interesting engineering is manipulation under precise positioning constraints, where AI planning meets classical control and the cost of an error is physical. It's off the software-agent core beat, but the constraint discussion (safety-critical, high-precision, real hardware) is a useful contrast to text-only agents.

Johann Rehberger found an attack against Claude Code's auto mode that works about 80% of the time: trick the agent into downloading and unzipping an archive, then run code that imports base64 without noticing a malicious local struct.py gets executed along the way. The sharpest detail is the failure mode — in several runs Claude detected the compromise and tried to kill the malware process, but auto mode blocked its own cleanup command. The safety layer became part of the exploit.