AI Agents

ORCA closes a real gap: automated program repair usually starts from issue reports or failing tests, but microservice incidents get diagnosed from telemetry. It distills paired failure/reference telemetry into a fault signature, uses that to localize candidate code and deployment-config, and generates unified-diff patches verified by telemetry replay that separates patch validity, semantic correctness, and test-oracle integrity. On a 575-case benchmark it beats baselines on cost-effectiveness. The telemetry-grounded verifier is the standout — it catches outcomes issue- or test-only evaluation would miss.

Agent Lightning v1.0 formalizes "harnessed agentic RL": the deploy-time harness owns the environment loop while the trainer only observes LLM request/response pairs, which breaks retokenization, sample merging, and advantage calculation in non-obvious ways. In ~3,500 lines it supports arbitrary harnesses and ships a reproducible coding-agent pipeline. The headline result — Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified with 6K examples and modest compute — is a 14.6-point gain that's worth reproducing. The disaggregated architecture has already influenced verl, AReaL, and slime.

A sharp negative result. The authors re-evaluate two memory-based self-improving agents along two axes prior work skipped: multiple runs to quantify variance, and shuffled task order. Both reveal fragility — agent evaluation is already noisy on multi-step tasks, and the self-improving loop amplifies it, while reported improvement depends heavily on task order that imposes a hidden curriculum. Adding rubrics and environment feedback to memory construction partially closes the gap but doesn't eliminate it. This is the kind of reproducibility audit the self-improvement literature needs.

StartupBench builds its tasks by reverse-engineering AI products with demonstrated market adoption, rather than researcher-selected tasks — a deliberate answer to the complaint that benchmarks measure the wrong work. Tasks are deliverable-oriented, graded with fine-grained rubrics under a unified harness. The result: even the strongest model completes only about 30%, despite substantial partial progress, with complex instruction-following and domain expertise as the main failure sources. The methodology — grounding task selection in adoption data — is the contribution worth borrowing.

An unusual cross-disciplinary study: across 10,000+ communities of LLM agents exchanging messages and revising opinions, group dynamics collapse into three regimes — indifference, polarization, consensus. On objective math questions communication improves accuracy; on subjective political statements it drifts groups rightward. The authors fit a statistical-mechanics model where agents favor lower social pressure, and it predicts individual trajectories from initial opinions, generalizes to unseen community graphs, and reproduces the archetype distributions — outperforming standard baselines. The predictive validation makes this more than an analogy.

A hands-on writeup of an almost-fully self-hosted, sandboxed agentic build pipeline — the kind of concrete infrastructure post that's rarer than the framework announcements. The value here is the operational detail: how untrusted agent-generated code gets isolated, how the build stages hang together, and where self-hosting hits friction. 104 points and 55 comments on HN suggest the tradeoffs resonated. Worth reading if you're designing a code-agent factory rather than wiring an API.

McAteer's thesis: the agent harness — the scaffolding that manages tools, context, and control flow — is steadily being absorbed into model weights, and what's left over becomes an interface for directing human attention rather than steering the model. It's a conceptual framing piece, not a technical writeup, but the observation lands: much of what harness engineering does today is transitional, and the durable design work shifts to where humans plug in.