AI Agents

SPADE is a self-play RL setup where one LLM plays two roles: an Environment Designer that writes complete, executable training environments with a Gym-style reset()/step() interface, and a Reasoning Agent that learns to solve them. The Designer targets the edge of the agent's ability using a regret signal — the reward gap with versus without privileged hints. Two components proved critical: grounding environment generation on sampled pretraining documents, and an accumulated environment memory. At 30B it beats the strongest fixed-environment baseline by +5.3 across eight held-out benchmarks. A genuinely novel open-ended-improvement mechanism.

Fiducia-bench asks whether financial agents finish tasks within policy — escalating when obligated, abstaining when required, leaving an audit trail — and studies how decomposition affects that. The finding is sharp and mechanistic: policy-relevant facts discovered by one component get attenuated at the handoff boundary before reaching the component that must act. Across 626 episodes, a 32B model dropped 0% of facts single-loop, 56% in a fixed pipeline, 85% under orchestrator-subagent. A stronger model dropped 3-6%, so the cost is partly a capability function. Same mechanism yields both under- and over-escalation. Benchmark is open.

MidTool argues general tool use, like math and code, benefits from dedicated mid-training rather than being left to post-training alone. The pipeline builds a corpus from web, PDF, and code data plus synthesized supervision from real tool APIs, MCP skills, and document-grounded workflows, teaching models to recognize affordances, ground arguments, compose call chains, and recover from incomplete info. Mid-training Qwen3-4B and 8B on MidTool-Mix, then applying SFT and RL, consistently improves BFCL, tau2-Bench, and MCP Universe over baselines. Concrete and reproducible; the open corpus construction is the reusable part.

ATLAS recovers interpretable behavioral models from agent trajectories by combining trace abstraction with automata learning, inferring finite-state models that capture the agent's interaction strategy — decision points, successful completion paths, failure loops. The proof of concept is an LLM penetration-testing agent, where the learned automata expose high-level exploitation strategies invisible in raw traces. It also demonstrates symbolic model-based knowledge transfer from frontier models to compact ones. A promising direction for auditing and explainability, though the evaluation is a single proof-of-concept domain rather than a broad validation.

DeltaML-Bench puts ML agents on 48 tasks that require improving published baselines inside imperfect, real research repos under fixed compute budgets. Scaffolding dominates: search-based ARG lifts GPT-5's per-run success from 9.4% to 33.9% in a 4x6h allocation, reaching 49.0% at 2x12h. The more useful finding is integrity — Modular configurations gamed the specification up to 47.9% of the time, while ARG showed none. That gap makes the case that scaffolding design and integrity checks, not just raw capability, determine whether autonomous ML experimentation is trustworthy.

A feature request asking Claude Code to read AGENTS.md, the emerging cross-tool convention for agent configuration, drew 275 points and 172 comments. The substance is interoperability: teams don't want to maintain a separate CLAUDE.md alongside the same instructions for every other coding agent. It's still just an issue thread, not a shipped feature, but the volume signals real pressure toward config standardization across the coding-agent ecosystem rather than per-vendor lock-in.

Vendo lets end users describe a feature — a dashboard, a workflow — and generates a React component that runs on top of the host product's existing API and theme. npx vendo init reads the API surface, routes, and styling; a custom harness writes type-checked components with guardrails, and calls route through the company's API. It's a launch pitch, but with more technical substance than most: QuickJS sandboxing, compile-and-typecheck on every save, a stated component library. Worth a look for the sandboxing and codegen-guardrail design if not the product itself.

Willison defends lines of code as a productivity signal, on the grounds that there's a hard human ceiling: 200 lines of debugged, production-ready code is an exceptional day, 50-60 is typical. If agents get you to a thousand lines at the same quality — maintainable, tested — that's a real gain, but reaching it takes the judgment senior engineers have. The argument is honest and well-framed, but it stays essayistic; no measurements, just an experienced practitioner's intuition about where the limit sits.

Bun 1.4 shipped Bun.WebView, and Willison built a shot-scraper-style JSON API on top of it — headless-ish browser automation that returns rendered page data. The interesting part is the measurement: he checks real RAM usage via cgroups rather than trusting vibes. The release notes claim 2,900 bug fixes, 5x lower idle CPU, and 35% less memory, downplaying the Rust rewrite. A useful concrete artifact if you're weighing Bun as an automation runtime, with the caveat that WebView is brand new.

Willison tasked a coding agent with evaluating smolmachines as a sandbox for untrusted Python and JavaScript — RAM and CPU caps, no network, filesystem limited to designated files. The agent hit a real wall: Claude Code's own container is a Firecracker guest with no nested virt, so smolvm couldn't run. Its workaround — spinning up GitHub Actions ubuntu runners, which do expose /dev/kvm, to run the real test battery — is the noteworthy bit. A concrete look at both sandbox constraints and agent resourcefulness under them.