AI Agents

A genuinely surprising negative result: after evolving a harness around a weaker model, training that model on a stronger expert's full trajectories under the same harness regresses performance by 4 to 30 points across all seven enterprise tasks — even though the identical procedure helps under the un-evolved harness. The diagnosis is model-harness fit: imitation imports the expert's planning style the weaker model can't execute. Their fix, an on-policy pipeline that rewrites only the failing turn, preserves native planning and keeps both gains.

FrogNano is a 4B coding agent post-trained purely with RL on ~1,500 SWE environments, no distillation from larger models. The claimed key ingredient is online task synthesis that generates problems calibrated to the current checkpoint's learnability frontier — hard enough to teach, easy enough to solve. The interesting bet is that a small model can reach competitive SWE performance without a frontier teacher. It's a technical report, so weigh the eval breadth against the absence of head-to-head distillation baselines.

SkillAdam treats agent-skill self-evolution as optimization over discrete, non-differentiable skill documents, borrowing Adam's structure: an optimization memory acts as a first-moment analogue to stabilize edit direction, and a volatility-driven edit budget acts as a second-moment analogue to size each revision. The motivation is sound — heuristic skill-update loops drift and overwrite prior corrections. It reports state-of-the-art across seven short- and long-horizon benchmarks with fewer iterations. The Adam analogy is more metaphor than proof, but the direction-stability failure it targets is one you've seen.

AttnCompress compresses SE-agent trajectories dynamically instead of static pruning: it segments on perplexity spikes to keep code and log structure intact, uses proxy attention weights to score how relevant each history block is to current reasoning, and re-recalls dropped context via a rolling window. On SWE-Bench-Verified and Multi-SWE-Bench it reports 53.17% pass rate while cutting tokens 21.6% and cost 33.6%, and claims model-agnostic behavior. The rare compression result that improves pass rate rather than trading it for savings.

A well-designed study of the production assumption that agents are interchangeable by role. Eight teams per setting form independently, keep private notebooks over ten episodes, then swap role-matched agents. Task scores barely move, but communication cost per unit of progress rises 16-63%, and in Hanabi a swapped agent costs more than an inexperienced one — evidence of interference from conventions learned with a former partner. The placebo control isolating roster-change disruption from who-fills-the-seat is what makes this credible.

A write-up on the VM infrastructure behind mobile coding agents like Claude Code and Instinct. This is the deployment layer most agent discussion skips: how you give an agent a real shell, filesystem, and network without handing it your host. Modest HN traction (75 points) and the detail lives in the linked post, but the isolation-plus-provisioning problem is exactly what you hit the moment coding agents leave the demo and run against real repos.

A small skill artifact targeting a specific coding-agent failure mode: burying the actual answer under preamble, caveats, and restated context. The framing is glib but the complaint is real — agents that pad output waste tokens and reader attention, and the fix is a reusable instruction rather than a one-off prompt. Strong HN interest (462 points) suggests the annoyance is widespread. Treat it as a pattern to steal, not a library to depend on.

Simon Willison's account of OpenAI's claimed Navier-Stokes resolution — reportedly produced by an unreleased model — is more useful as an incident report than a math result. The interesting details for practitioners are the scale (per the linked coverage, roughly 10,000 agents and 130B tokens, north of $40M) and the priority dispute with NYU's Buckmaster and Anthropic's Alpöge, who used Claude and Codex over nearly a year before an August 15th breakthrough. Narrative, not reusable technique, but the swarm economics are real.