AI Agents

A training-free way to shorten reasoning chains: a small model reads the big model's chain-of-thought twice — once prompted to be brief, once to be verbose — and the difference between its two logit vectors is added to the large model's logits at every step. On seven math/science/coding benchmarks it cuts Qwen3-32B's CoT by 31% with unchanged accuracy, and trims a self-debugging coding agent's per-turn tokens by 20%. Ablations show most savings come from killing repeated checking after the answer appears.

Instead of continuing a long-horizon agent from a compressed interaction history, CarryOn organizes context around reusable task objects — continuation carriers plus complementary execution context plus a live tail of recent turns. The hard part is keeping those carrier records valid as old history is retired, which it handles by revising affected records against bound evidence rather than rebuilding everything each fold. On OSWorld 2.0 and BrowseComp-Plus it gains 5.8% and 7.0% while cutting total tokens by 51.9% and 22.9%.

A clean impossibility result for safe self-improvement. Model agents as programs and define improvement via a semantic score that needn't be computable; then show no terminating gate — certifier or search procedure — can be both reliable (never approve a non-improvement) and non-conservative (never miss a real one). Self-referential constructions force failure, and randomized gates with bounded error hit the same wall. The practical escape is the decidable regime already used in practice: bounded evaluation plus a finite proposal budget.

A runtime framework that keeps a coding agent's project state explicit without touching model weights: the agent records its current problem, key findings, discovered repo relationships, and unfinished work, while runtime support searches for affected callers, tests, and docs and maintains a compact working brief. A separate diagnostic session revisits history when behavioral signals accumulate. On RefactorBench full-chain completion rises 29.4%→35.3% and cleanup accuracy 87.0%→95.4%; LHTB mean reward 0.427→0.576 at lower execution cost; ChainSWE bug-fix 35.1%→39.5%.

A source-aware taxonomy for why agentic workflows fail, built from literature plus GitHub evidence. The core structure is a two-axis matrix — workflow stages on one axis, agent capabilities on the other — that localizes intermediate process states rather than just classifying final symptoms. The empirical split is the useful part: papers emphasize planning, reasoning, and perception failures, while GitHub issues concentrate on execution, tool, API, and infrastructure failures, with iterative execution and actuation the dominant hotspot. Reliability is a systems problem, not a model-output problem.

A FOSS hook that intercepts your push-to-prod step and makes you answer a questionnaire about how the AI-generated code you're shipping actually works. The guardrail is social, not technical: it forces a comprehension check rather than scanning for bugs. The pattern is interesting — comprehension debt is real when agents write most of the diff — but the Show HN is thin on how questions are generated, how answers are graded, or how to stop it becoming a rubber-stamp.