01 Anthropic’s agent marketplace executed live transactions, one linking a Mill Valley estate to equity
Anthropic built and ran a test classified marketplace where autonomous agents represented both buyers and sellers and struck deals that involved real goods and real money. The company used the experiment to see how agent‑to‑agent commerce might function when machines negotiate, list, and close transactions without a human intermediary driving each step.
One notable listing in the test offers a 13‑acre property in Mill Valley, north of San Francisco, under an unusual payment condition: the seller is asking for Anthropic equity as consideration. The listing shows that agents can negotiate nonstandard terms and facilitate transfers that mix financial instruments and physical assets.
The experiment demonstrates practical capabilities — agents can create listings, evaluate offers, and transfer value — but it also surfaces immediate governance and operational questions. Who verifies identity and title when an agent signs a contract? How are payments, escrow, and regulatory checks integrated? The test does not resolve these issues, but it makes them concrete.
If agent marketplaces scale, commerce could get faster and more automated, particularly for complex or bespoke deals. At the same time, companies and regulators will need new rules for accountability, billing practices, and dispute resolution when autonomous systems act as counterparties.
- Anthropic ran a live test marketplace where agents represented buyers and sellers and completed deals involving real money and goods.
- A 13‑acre Mill Valley property appeared in the marketplace with a request for Anthropic equity as payment.
- The experiment highlights operational gaps: identity verification, escrow, and legal responsibility are unresolved in agent‑led transactions.
What moved around the edges
Agent memory that mimics human forgetting hits 52% Recall@5 and cuts token waste
A GitHub project applies the Ebbinghaus forgetting curve to a vector-store memory, assigning strength scores that are reinforced on recall and pruning unused entries; on the LoCoMo benchmark it reports 52% Recall@5 and an estimated ~84% reduction in token waste, and it layers a graph over the vector store and runs as a local‑first MCP server using DuckDB.
Hacker News AISusHi Tech Tokyo 2026 turns Tokyo into a demo‑first show with four dedicated domain floors
SusHi Tech Tokyo 2026 organizes its program into four tightly defined technology domains, each supported by live demonstrations, dedicated exhibit floors, and sessions that bring together builders and funders from around the world.
TechCrunch AIStudy identifies the mysterious golden orb found two miles beneath Alaskan waters
Researchers have published a study that resolves the identity of a bizarre golden object discovered roughly two miles under the Alaskan seafloor, settling questions that followed the initial recovery.
404 Media AIDeepSeek launches V4 Pro (1.6T‑A49B) and Flash (284B‑A13B) with Ascend support while Tiger loses top‑bench spot
New DeepSeek releases include V4 Pro (1.6T‑A49B) and Flash (284B‑A13B) in both base and instruct variants and are reported runnable on Huawei Ascend chips; reporting also notes the ‘Tiger’ model family is no longer the benchmark leader.
Latent SpaceStudent seeks collaborators to build an open, local‑first NotebookLM‑style research assistant
A college sophomore launched an open project to create a local alternative to proprietary research assistants—aiming to convert documents into reusable, researchable assets—and published the code on GitHub while inviting collaborators to contribute.
Hacker News AI