Design the Handoff Before the Agent

Anthropic said Claude now leads 26% of its AI R&D work.

That is the loud number.

The quieter one is more useful: more than 90% of the measured work now involves AI collaboration or AI leadership, with no category running fully autonomous.

That should kill two bad stories at once.

The first is the magic-agent story, where software quietly replaces whole functions if the model gets strong enough. The second is the harmless-copilot story, where AI sits politely beside the real workflow and waits to be asked.

Anthropic is showing the middle layer: shared work, supervised authority, explicit boundaries.

That is where normal companies need to look. Not at whether an agent can do everything alone. At which parts of the work it can carry, where a human still owns judgment, and what gets reviewed before the output becomes real.

The handoff is the product.

If your team cannot name the handoff, it is not ready to scale the agent.