A new framework lets an AI agent patch a broken plan instead of scrapping it and starting over.
Plan-and-Patch pairs a diffusion language model with a plan-and-act loop: the model writes a structured, program-like plan in one parallel pass, then repairs only the parts invalidated by a failed tool call or bad assumption, leaving the rest of the plan untouched. The researchers tested this against a standard autoregressive model, comparing an 8-billion-parameter diffusion planner (DreamReasoner-8B) to an 8-billion-parameter autoregressive one (Qwen3-8B). On the Natural Plan benchmark, with no extra training, the diffusion model fixed broken plans correctly 53.7% of the time versus 27.0% for the autoregressive model, nearly double. After both models got additional training on two agentic benchmarks, ALFWorld and TextCraft, they produced working plans at similar rates, but the diffusion model did it 39-46% faster.
That speed and repair gap matters because long-horizon agents spend most of their time recovering from surprises, not executing a perfect plan. Regenerating an entire plan from scratch every time a tool returns something unexpected is slow and risks introducing new mistakes elsewhere. Editing just the broken section, the way a patch works in version control, is a more targeted fix.
The results come from two simulated environments built for benchmarking, not messy real-world deployments, so it is still an open question whether the patching trick holds up once agents are juggling actual APIs and actual failures.