In today’s Digest, we cover Pinterest launching CTV audiences following its tvScientific deal, Google signing a Pentagon agreement for ‘any lawful’ AI use, and Anthropic testing AI-to-AI commerce in an internal marketplace pilot.

Pinterest launches CTV audiences 

Pinterest is accelerating its move into connected TV (CTV) following its acquisition of tvScientific, bringing its audience data into the performance TV ecosystem. The rollout allows advertisers to access Pinterest’s high-intent audiences directly within CTV campaigns.

The company is betting on shifting viewing habits, with CTV ad spend in the US expected to overtake traditional linear TV in the coming years. By combining its intent-driven user signals with tvScientific’s measurement and buying capabilities, Pinterest claims advertisers can drive stronger performance outcomes, including higher purchase rates. 

Google signs Pentagon deal 

Google has reportedly entered into a classified agreement with the US Department of Defense allowing its AI models to be used for ‘any lawful government purpose’. The deal places Google alongside OpenAI and xAI, which have also secured government AI contracts, while Anthropic was previously excluded after refusing to loosen safeguards tied to weapons and surveillance use. 

Details of the agreement suggest limits on applications such as domestic mass surveillance and autonomous weapons without human oversight. However, the contract reportedly does not grant Google authority to veto how its technology is ultimately deployed, raising questions about how enforceable those safeguards are in practice. The deal also includes provisions requiring Google to adjust safety settings at the government’s request.

Anthropic builds AI agent marketplace 

Anthropic has trialled a closed marketplace where AI agents acted as both buyers and sellers, autonomously negotiating transactions involving real goods and money. The experiment, dubbed Project Deal, involved 69 employees who were each given USD$100 (£79) in gift cards to participate. Despite its limited scope, the test generated 186 transactions worth more than USD$4,000 (£3,160), with the company noting it was struck by how well the system functioned.

Anthropic said it operated four parallel versions of the marketplace using different AI models. One was a ‘real’ environment, where participants were represented by its most advanced model and transactions were fully honoured after the experiment, while the remaining three were controlled test environments designed to study how varying model capabilities influenced outcomes. 

The pilot also highlighted emerging dynamics in agent-driven economies. Anthropic found that more advanced AI models consistently secured better outcomes in negotiations, even though participants were largely unaware of the disparity raising concerns about potential agent quality gaps. The company also observed that initial instructions had little impact on deal outcomes, suggesting that model capability, rather than prompt design, may play a more decisive role in autonomous commercial interactions.