The Agent Said It Was Done. The Database Disagreed.

Researchers are turning every layer of the agent stack—skills, harness, model, environment, and orchestration—into something that can optimize itself from execution traces instead of manual tuning.

• Microsoft's SkillOpt automatically edits skill text files based on scored trajectories, only accepting changes that improve a held-out validation set.
• Google Research's WikiSkill builds a structured 'wiki' of agent successes/failures to prevent repeating rejected fixes when optimizing skills.
• Self-Harness mines execution traces to rewrite harness code (tool use, context, control flow) and showed relative gains up to 132% on Terminal-Bench-2.0, SWE-bench Verified, and AppWorld.
• Sakana-style Darwin Gödel Machine lets a coding agent modify its own implementation and keep an archive of variants, while Meta's Hyperagents extends self-modifying harnesses beyond pure coding tasks.
• Xiaomi's HarnessX enables model-harness co-evolution (open-weight models only): harness-discovered strategies become fine-tuning data, which then improves harness optimization in a loop.
• EnvHarness evolves the training environment itself (not just the agent), yielding up to 9-point gains with 9.8% fewer interaction steps; EverMind AI's Raven applies self-evolution to multi-agent orchestration.

Runway is pivoting from Hollywood-only video gen to agentic ads and robotics, betting that video-model scaling laws generalize across domains the way LLM scaling did.

• Runway launched 'Runway Ads,' an end-to-end agent that creates, edits, publishes and tracks performance of paid ad campaigns — not just generates clips.
• Runway released Praxis-1, an open-weight 'world action model' that adapts to new robot embodiments with light fine-tuning, extending video-gen tech into physical/robotic AI.
• CEO Cristóbal Valenzuela says Hollywood studios have moved from 'denial' to full 'acceptance' of AI video over the past ~18 months — the debate is now about scaling adoption, not legitimacy.
• Runway's thesis: video models do 'next-frame prediction' analogous to LLM next-token prediction, so the same scaling laws apply — explicitly betting against the 'LLMs have plateaued' narrative some labs push.
• Runway sees 12-18 months to production-ready physical AI deployments; data isn't the bottleneck, but cleaning/annotation pipelines are the real constraint.
• Valenzuela claims China is 'world-model-pilled' (70% of token consumption is video, heavy robotics investment) while the US remains 'LLM-pilled' — a structural gap he thinks Western labs need to close.

We're going to need default hard budget caps on pretty much everything

• <p>Here's a product feature which the world is going to need a whole lot more of over the coming months and years: <strong>default hard budget caps</strong>. I'm talking about the feature of pay-by-usage services and APIs that lets you say "after $X/month, cut this thing off and return errors". Thes