The headline improvement in Claude Sonnet 4.7 is sustained tool use. It strings together more tool calls before it needs you, recovers from a failed call instead of stalling on it, and holds a longer workflow in view without losing the thread. In practice that means an agent that used to hand control back every few steps now runs the whole sequence and reports when it's done.
That's a real capability gain. It's also the moment identity stops being optional.
The checkpoint was doing two jobs
When a human approved every step of an agent's workflow, they were doing something besides clicking. They were the identity check. A person confirming a charge is a person you can name and hold accountable. Every capability improvement that removes a human checkpoint also removes that — quietly.
Sonnet 4.7 is good enough that the person now approves a goal and walks away. "Procure the vendor renewals, keep us under budget." Twelve tool calls later something got bought, and the human who used to be the accountable party in the loop was never in the loop.
The human step you just automated away was also your proof of who authorized the action. Remove it and you have to put identity back explicitly — or you have an autonomous system spending money on behalf of no one you can name.
Putting identity back — once, at the action
The fix isn't to slow the agent back down. It's to verify the operator behind it at the point where it matters — the moment it acts on money — instead of at every step it no longer stops on.
1// One check at the payment action, not twelve checks the agent no longer pauses for.
2const result = await agentscore.assess(walletAddress, { policy });
3
4if (result.decision === 'deny') {
5 return res.status(403).json({ reasons: result.decision_reasons, verify_url: result.verify_url });
6}
7// Verified operator behind the autonomous run → proceed.A procurement run
A Sonnet 4.7 agent works a multi-step procurement: compares vendors, picks one, assembles the order, pays. The identity check fires once — at the pay step — resolving the operator to a KYC-verified, sanctions-clear entity. The other eleven steps stay fast and unattended, because they should. You gated the one that moves money, not the agent's autonomy.
The point
Better models make agents more useful and more autonomous in the same release. The second half is the part that changes your risk: the more of the workflow the model can carry alone, the more the "who is behind this" question is one you have to answer on purpose. Sonnet 4.7 is a good reason to answer it now.
Shipping autonomous agents?
See how identity verification worksReady to power your agents with secure commerce?
Join innovators using AgentScore to accept payments, verify buyers, and ensure compliance for every AI-driven transaction.
Read More Blog Posts
Anthropic's Claude Fable 5: Mythos-Class Speed for Million-Token Workflows
Building agents that reason over massive codebases and run for days requires models that maintain focus and autonomy at scale. Anthropic's Claude Fable 5 brings
Gate AI agent escrow payments with credential verification
AI agents accessing real estate escrow systems risk fraud through impersonation and credential spoofing without verification. AgentScore enables merchants to ga