AI Agents

A large-scale diagnostic (81,390 replayed candidate pools across five benchmarks) isolating where multi-agent systems lose correct answers. The core finding: a correct answer is often already among the generated candidates, but consensus-without-quality-control converges on popular errors — the authors call it memetic drift. Judge reliability isn't a model trait; it varies with task, generator, and answer rarity. Combining answer frequency with judge signals lifted accuracy from 63.8% to ~71%, mostly by rescuing outnumbered correct answers.

Prior self-improving agents cap out at roughly meta-depth two because editing your own editing machinery is destabilizing. Meta^n sidesteps this by keeping the meta-operation fixed and recursing on its growing input instead — each layer reads the traces and code below it, then writes the next layer as a pre-process plus callable helpers. Depth is set by convergence, not fixed in advance. It beats prior self-improvers across eight benchmark families and is the only method scoring above zero on ARC-AGI-2.

A controlled ablation on a production scientific-agent platform, using a verifiable protein-function task, disentangles what drives quality. Model choice dominated: Opus at 92-94% versus o4-mini at 40-50%. But a PPO policy hit 88% at zero token cost, fastest latency, and perfect consistency — with no reasoning trace. Federation topology imposed negligible penalty. The guidance is blunt: for routine verifiable tasks, a cheap deterministic policy is near-frontier and fully reproducible; reserve flexible LLM reasoning for open-ended discovery.

StarHarness evolves the agent harness — prompts, tool interfaces, skills, MCP providers, subagent structure, loop config — while keeping model weights frozen. It stratifies tasks by baseline failure behavior, separates proposer-visible search tasks from hidden selection tasks, and holds out generalization tasks. Across ITBench SRE, EnterpriseOps ITSM, and AutomationBench Finance, it gains 20-35 points over the default harness after only 4-12 accepted changes, and those gains transfer across GPT and Qwen families without re-evolution. Trace analysis credits interface repairs and encoded operational conventions.

OODA-Tool argues that ReAct-style policies suffer state-action competition: interleaving state tracking and action generation in one autoregressive trajectory lets the pressure to emit the next call overwrite accumulated context. Borrowing Boyd's OODA loop, it routes each decision through controller-checked typed states — Observe reconstructs state, Orient decides whether to act, Decide forms an admissible action, Act emits it. Tested on Qwen3 from 0.6B to 14B, it consistently improves multi-turn success, with the largest gains on smaller models and history-dependent tasks.

Warp's writeup on building code agents that improve from their own usage on Claude. The production angle is the draw — how they close the loop from real developer sessions back into agent behavior — but as a vendor post it's likely thin on the API-level mechanics you'd want to replicate. Modest HN traction (58 points). Read it for the pipeline shape and eval framing, not for a reproducible recipe.

Applies domain-driven design — bounded contexts, aggregates, ubiquitous language — to agent architecture, arguing that agent boundaries should follow domain boundaries rather than tool inventories. It's a pattern piece, so its value depends on how concretely it maps DDD constructs onto planning and tool decomposition versus just borrowing vocabulary. If you're already structuring multi-agent systems and reaching for organizing principles, this is a useful lens to test against your own decomposition.