Google Sparrow-2: On-the-Fly Contextual Planning in Large Language Models

LLMs have always been great at completion, but mediocre at planning—lining up actions, revising steps, and iterating based on feedback. With Sparrow-2, Google is closing that gap: the model can now construct multi-step plans, evaluate their feasibility, and self-correct within a single conversation context. It's a big deal for engineers building agents that need more than just 'do what I say.'

How does it work?

Sparrow-2 uses a new recurrent attention mechanism that lets the model 'slot in' plans, subgoals, and results as named objects in its working memory—think of it like an internal scratchpad that persists throughout a session. The key innovation is that the model can dynamically revise those plans based on user feedback or new data, without losing context or starting over. This feels much closer to how humans actually reason—draft, revise, adapt.

Why it matters for engineers

If you're building multi-agent workflows, scriptable LLM APIs, or anything that requires agentic reasoning, this is a game-changer. No more hacks with toolformer plugins or out-of-context chains. You can now ask, "Plan steps to automate this workflow, handle exceptions, and retry if needed"—and Sparrow-2 can keep it all in context, updating as things change. It means less glue code, more robust automation, and, honestly, more trust in your LLMs.

I expect this to show up soon in Google Cloud's Duet AI and Workspace automation hooks. If your app lives or dies by agent reliability, start poking at this API now.