If you're building any agent that outputs structured artifacts, this is the failure loop you should be running. paper: Limitations: It would be pretty lengthy to improve on sparse graphs

what I'm taking away: if you're serving large models and already using speculative decoding, Saguaro is the next step and move the draft to a separate device, add a speculation cache, parallelize.…

what I'm taking away: If you're running debate loops, planning agents or critic-refiner pipelines Your agents are probably already suffering from this. Add attention steering between rounds Score the…

what I'm taking away: if you're using GPT-5.4 or Gemini to audit your agent's actions, there's a 4B open model that's more accurate, runs locally and costs a fraction plug it in as a trajectory-level…

what I'm taking away: if you're building code-as-action agents, the compiler is a free safety layer you're probably not using you don't need a sandbox, a policy engine, or a bespoke retry classifier.…

my biggest takeaway: skill persistence matters way more than people think if agents can remember, evaluate, and refine how they work, the improvement compounds faster than us finding optimisation…