Anthropic ships an agent that asks, and a bill for a judge

Anthropic opened a US research preview of Claude Duet, and what it led with is a list of what the agent will not do by itself. Duet runs on Opus 5.5 in its own app rather than inside Claude, keeps working while the user is away, and stops to ask before sending messages, making payments, signing documents or deleting files, per CNBC. The autonomy is the capability. The pause is the product.

Set that against this ledger's record of the same company and it reads as a turn. Anthropic shipped Claude Sonnet 5 in June as its most agentic mid-tier model, built to plan, drive browsers and terminals, and run autonomously (Anthropic), then Claude Opus 5 in July, aimed at long-running agents and professional work (Anthropic). The models were built to keep going without a person. The first app built on them is sold on four places where it will not.

Four rounds, one question

Today's funding asks that question in four industries: where does a human have to look?

• Vessla left stealth with a $3.3 million pre-seed led by Seedcamp, with Inception Fund and Tiny VC joining, for agents that run property management workflows and hand decisions needing human judgment back to staff (Tech.eu). The handback is the pitch, funded by a firm that writes roughly $1 million first cheques (TechCrunch).
• Relace raised £1.2 million in pre-seed led by Boost Capital for software that works out which financial regulations apply to a business and flags where it is falling short (Dealroom). It files nothing. It decides what a compliance officer reads.
• Greptile raised a $50 million Series B led by Menlo Ventures at a $750 million valuation and is pushing beyond AI code review into testing and runtime validation (Runtime). That moves the check from before the merge to after it.
• Reduck raised $3 million in pre-seed led by Daphni for web automation that lets AI agents act on sites without going through traditional APIs (Tech.eu). An API is a gate someone built on purpose, and this round is capital for going around it.

Three of the four put a person back in the loop at a chosen point. The fourth is funded to remove the gate the web already had.

The stop button is a courtroom

The same day, the same company went further. Anthropic privately circulated a draft federal bill that would let the Commerce Department ask a judge in the District Court for the District of Columbia to suspend a frontier model, but only on a foreseeable, imminent risk of more than $1bn in damages or the death or serious injury of more than 50 people, while barring every other federal agency from suspending a model at all (TechCrunch).

That is Duet's approval dialog written as a statute: one named approver, a short list of qualifying actions, and a threshold below which nothing interrupts anything. Whether $1bn and 50 lives is the right line is a policy argument worth having. The shape of the answer is not an accident.

What runs while nobody is watching

The plumbing is being built for the unattended case. Red Hat, IBM, Google Cloud, CoreWeave and Nvidia's llm-d reached 1.0 as a CNCF-governed distributed inference layer, with session-aware routing for agentic workloads and router high availability with rolling updates (Red Hat). Sessions that have to outlive a deploy are a requirement written for something that does not stop.

So the gate is becoming the specification. Duet's four actions, Vessla's handback, Relace's flag and the bill's two numbers are all answers to the same design question, and of today's rounds only Reduck's is funded to answer it the other way. The figure worth tracking is not what the models can do unattended: the record already says they were built for it. It is how short the approval list gets with each release.