VCs are backing agents that prove an outcome, not agents that promise a platform

The data loop thesis

Andreessen Horowitz framing of Databricks is the clearest version of the data loop thesis. The firm says Databricks has surpassed a $7 billion revenue run rate, is growing at 80% year-over-year, sees more database queries from agents than from humans, and agents launch more than 17 million databases on Databricks every day. Lakebase, a Databricks product, has crossed a $100 million revenue run rate, according to the same write-up. Those numbers are reported by a venture firm about a portfolio-relevant company, so treat them as investor framing rather than audited results. Even with that caveat, the direction is the point: if agents already query more than humans, then owning the system where those queries run is a different kind of moat from owning a model. m:…1875

The mechanism behind the thesis is grounding. Databricks acquired Neon, a serverless Postgres company, in 2025, according to the write-up, and grounding an agent in a company ontology raised answer accuracy from 50% to 84.5% in the example given. The role of human workers is shifting from performing tasks to specifying, delegating and judging agent work, the firm says, and sustainable competitive advantage in an agentic firm requires owning the data and decision loops. That last claim is the investment test in one sentence: a startup that rents its data and its decision loop from someone else is renting its moat. A startup that owns the loop, and can show accuracy improving as the loop turns, has something a competitor cannot copy by swapping a model. m:…1875

Fraud defence as an agent market

Cowboy Ventures thesis starts from the other side of the same coin: agents are also the attack. The firm says AI-enabled fraud is estimated to be 4.5x more profitable than human-powered methods, the annual cost of fraud is approximately $150 billion, representing a 5x increase since 2020, and Americans over 60 reported $7.7 billion in losses to the FBI in 2025. In a 2024 Harvard study cited by Cowboy Ventures, AI-generated spear-phishing emails achieved a 54% click rate, while generic phishing emails and those written by human experts achieved a 12% click rate. People over 70 hold 31% of US household wealth, according to the firm, which explains why the target group matters: the most vulnerable group holds a large share of the money. m:…8337

Cowboy Ventures says it is seeking to invest in multi-modal scam defense solutions, pooled scam intelligence and payment interception solutions, agent-ready secrets management solutions, impersonation takedown tools for consumers and small businesses, and scam-loss protection bundled with detection. Personal agents like Muse and Instinct are susceptible to scams, according to the firm, and impersonation is the most common scam type, with scam infrastructure productized and sold as a service. The thesis is therefore not a single product but a stack: detect the scam, intercept the payment, take down the impersonation, and protect the loss. For a founder, the useful signal is that the firm is looking for a defence stack, not a single detector, because a single detector is easy for a productized scam operation to route around. m:…8337

A seed and a pilot with numbers

Playground Global seed round for XRlabs shows what a provable workflow looks like in a regulated setting. The firm is leading an $8.4 million seed round for XRlabs, according to Playground Global, and XRlabs trains its AI models on anatomy, safe access points, tissue and pathology. XRlabs SIM converts patient CT and MRI scans into interactive 3D models for surgical planning, and the platform has been used in more than 100 cases on three continents. XRlabs achieved the first-in-human use of physical AI in cranial neurosurgery at Lenox Hill Hospital, according to the write-up, and the system there allows surgeons to reframe and refocus the exoscope using gaze and voice. m:…3050

8VC write-up on BigHat and Anthropic gives a lab version of the same pattern. Anthropic and BigHat Biosciences announced the results of their antibody optimization pilot, according to 8VC, and the pilot involved an Anthropic researcher and Claude using BigHat to optimize two antibodies over seven rounds. Optimized antibody variants bound targets up to nine times tighter, remained folded up to 17°C higher than the parental antibody, and Claude processed 1,232 designs during the pilot. Antibodies are the single largest pharma category by revenue, according to 8VC, and the firm says the technical bottleneck in drug discovery has shifted from models to data infrastructure, with the research focal point shifting from compound libraries to labs that generate and validate molecules. A pilot with a tighter binder and a higher folding temperature is a result a lab can verify, which is why it travels better than a benchmark score. m:…0485

What the theses share

This suggests a practical filter for a VC or a founder reading this edition. Ask what improves as the system is used: does accuracy rise because the data loop turns, does fraud loss fall because the defence stack learns, does surgical perception improve because cases accumulate, does a binder get tighter because the lab validates designs. Sustainable competitive advantage in an agentic firm requires owning the data and decision loops, as Andreessen Horowitz puts it, and the other theses in this story are versions of the same test in different markets. m:…1875 m:…8337 m:…3050