ClaudeDevs Effort Routing for Agent Tasks

ClaudeDevs, the team that tinkers with how Anthropic's models get deployed in the real world, has been kicking around an idea that sounds boring on paper but matters a lot in practice: effort routing. The pitch is simple. Not every task an AI agent handles deserves the same amount of computational muscle, so why treat a one-line summary request the same way you'd treat a multi-step coding job with real consequences if it goes wrong?

The concept works by sorting tasks according to risk before deciding how much reasoning, checking, or caution Claude applies. A low-stakes request — draft me a tweet, reformat this list — gets a light touch. Something higher-stakes, like generating code that touches production systems or making a decision with financial weight, gets routed toward a heavier, more deliberate mode where the model slows down and double-checks itself.

This isn't just an efficiency play, though it's clearly that too. Agentic AI has a cost problem nobody likes to talk about out loud: running a model at maximum reasoning depth on every single query is expensive, and it's often unnecessary. Effort routing is essentially triage, the same logic an ER nurse uses to decide who needs immediate attention and who can wait in the lobby. Apply it to AI agents and you get a system that spends its compute budget where it actually buys you something — accuracy on hard problems — instead of burning it uniformly.

There's also a safety angle buried in here. If a system can flag a task as high-risk and respond by dialing up scrutiny, that's a meaningfully different posture than treating every prompt identically. Anthropic has spent years talking about calibrated caution in its models; effort routing looks like an attempt to build that calibration into the plumbing of how agents actually operate, not just into the model's training.

What's still unclear is how ClaudeDevs plans to classify risk in the first place. Risk assessment is messy and subjective, and getting it wrong in either direction — treating something risky as routine, or something routine as risky — undercuts the whole idea. For now this reads more like an architecture proposal than a shipped feature, and the real test will be whether it holds up once agents are routing thousands of tasks a day with actual stakes attached.