AI Agents

CANOPY challenges the belief that outcome-only RL hits a ceiling on small open models for long-horizon tasks. It diagnoses two concrete failure causes — signal starvation (sparse rewards give gradient only when a rollout group mixes wins and losses) and policy drift (over-updating a small task pool collapses the sampling distribution) — and fixes both by scaling same-task exploration, keeping updates on-policy and KL-anchored to the agent's own action tokens. A Qwen3-14B policy topped the AppWorld leaderboard without SFT, skill libraries, or scaffolding.

This paper argues that aggregate accuracy on multi-turn tool-calling benchmarks hides the real problem. It decomposes failures into action-class miscalibration (choosing wrong among TOOL_CALL/ASK/REFUSE/CONFIRM) and execution failure, using a self-revealing bound Acc ≤ Gold Action Recall to expose when state graders mask miscalibration. The sharpest result: a single context-only perturbation moves accuracy in opposite directions across model families, up to +11.5 versus -21.0pp on the same scenario. Heavily tool-trained models look strong precisely where the diagnostic separates them.

ContextPipe observes that context assembly — deciding what goes in each prompt, in what order, and when to compact under a hard window and byte-sensitive cache — is structurally the same problem as relational query execution: hard budget, tiered cache, statistics. It borrows the discipline directly with a five-phase Plan/Bind/Optimize/Execute/Feedback pipeline, a cache-aware optimizer, and an EXPLAIN ANALYZE trace. On the SWE-bench Pro Qutebrowser subset it cut tokens 31%, LLM calls 23%, and response time 9% versus append-only, at the cost of lower KV cache-hit ratio.

Defense-as-Skill tackles a real gap: skill-augmented agents load reusable skills as persistent context, giving a malicious skill a durable channel to steer future actions once task and workspace state make an unsafe action look useful — so pre-install vetting is insufficient. The fix implements the guard itself as an installable, inspectable skill (SkillSonar) that checks sensitive actions against the user's task boundary. Evolved via MCTS on their SCOPE-R dataset, it cut attack success on GLM-5 from 0.482 to 0.104 (ID) and 0.606 to 0.115 (OOD).

A live trace model for long-horizon agents serving two consumers whose needs diverge: the human monitoring the run and the agent whose bounded context must absorb the trace. It's an append-only event ledger folded incrementally into typed run state, compiled into per-consumer views. Monitoring questions answered from the compiled view used ~14-15x fewer tokens at 5-7x lower cost with higher accuracy (0.85-0.87 vs 0.48). Caveat: questions were co-designed with the schema, so the authors label those results conditional on schema coverage.

A design proposal treating agent memory as an explicit file format rather than an opaque vector store or framework-internal blob. The pitch is portability and inspectability: memory you can diff, version, and move between runtimes. It drew solid HN engagement (190 points, 93 comments), and the framing is directly relevant to anyone tired of memory being an undocumented side effect of whatever orchestration library they picked.

Baseten maps the efficient frontier of LLM inference — the tradeoff surface between throughput, latency, and cost per token — as concrete engineering rather than a benchmark leaderboard. For teams running agents at volume, this is the useful framing: where you sit on the curve determines whether latency-sensitive interactive loops and batch agent runs should share infrastructure at all. Vendor-authored, so read the assumptions, but the frontier framing itself is sound.

DoltLite is a SQLite fork with Git-style version control, and the notable claim is process: it was built with roughly 2,000 agent-generated PRs. That makes it a real artifact for judging whether agent software factories produce maintainable systems code, not a toy. The engineering target — versioned SQLite semantics — is demanding enough that shortcuts show up fast, which makes the codebase worth inspecting as evidence about agent-driven development at scale.

Anthropic released Claude Fable 5.1 and Mythos 5.1 with a published system card and API changes — the launch that dominated HN (1266 points). The headline economics are the 75% prompt-cache price cut paired with ~70% more output tokens, which reshapes cost models for agent loops that lean on caching. Light on agent-specific mechanism in the announcement itself; the system card is where the safety and behavior details live.

The Anthropic news-of-the-day wrapper covers the Fable/Mythos 5.1 launch: new SOTA claims, a 75% cache price cut offset by 70% more output tokens, and the broader pattern of top OSS projects replacing drive-by PRs with agent-run software factories. The text is teaser-only, so the interesting operational detail — how the price/token math nets out for your workloads — lives in the linked material rather than here.

Paint.NET's author describes shipping a from-scratch, clean-room reverse-engineered Direct2D implementation for WINE — 180,000 lines, mostly "vibe coded" by Claude, unreviewed by choice because reviewing it is infeasible against a 700k-line codebase built over 20 years. The candid failure modes are the value: Claude repeatedly skipped the COM equivalent of AddRef() on reference-counted objects and produced bad architecture that needed babysitting. A rare honest account of what unreviewed agent-generated systems code actually looks like.