The $110M bet is not on a smarter model. It is on agents that work
Noeri raised $110 million to build the harness around agentic systems, and Bedrock Energy raised $24 million for AI-driven drilling. The money is going to the layer that makes agents work, not to the next model.
- Noeri raised $110 million in a fundraise led by Menlo Ventures, with Madrona investing alongside strategic investors Cisco Investments, NVIDIA and Snowflake. Self-reported m:…3471
- Noeri's approach combines test-time training, synthetic data for edge cases, and self-distillation, in which a model improves by learning from a better-informed version of itself. Self-reported m:…3471
- On the 20VC podcast, Jev creator Diogo Almeida said reliability is the largest bottleneck in AI, and that data matters more than compute. m:…6725
- Bedrock Energy raised $24 million for AI-enabled geothermal drilling. Self-reported m:…2591
The raise
Madrona announced that it is investing in Noeri's $110 million fundraise, led by Menlo Ventures. Strategic investors Cisco Investments, NVIDIA and Snowflake are participating. Noeri is building what Madrona calls a continuously improving harness for agentic systems, the layer that sits around a model and keeps it working on a real task. The approach combines test-time training, which lets a model keep updating on the specific problem in front of it, with synthetic data for edge cases that rarely show up in production and self-distillation, in which a model improves by learning from a better-informed version of itself. The company was started by Carlos Guestrin and Yejin Choi, who are co-founders and co-CEOs. Madrona points to early evidence: using an open model, at a cost of a few hundred dollars per problem, Noeri's approach set new state-of-the-art results on most of the problems it attempted across mathematics, GPU kernel engineering, algorithm design and biology. m:…3471
Why now: the bottleneck moved
The raise lands on a complaint that keeps surfacing in investor conversations: the models are good enough, and the systems around them are not. On the 20VC podcast, Jev creator Diogo Almeida said reliability is the largest bottleneck for AI, and that data matters more than compute. On the a16z podcast, Alex Imas and Tyler Cowen made a related point, that progress is accelerating and the faster things move, the sooner they hit the real bottlenecks. Their example was agentic swarms: unpack the idea, they said, and agents are still very bad at coordinating, mostly just doing parallel processing. That is the gap Noeri is selling into. A harness that trains at test time and learns from its own better-informed version is a bet that the next gains come from the system around the model, not from a bigger model. m:…6725 m:…1459
The same money, applied
The same pattern shows up outside software. Bedrock Energy raised $24 million to deliver on-site geothermal heating and cooling for enterprise real estate. The company says its drilling automation uses edge physical AI and downhole sensor fusion across more than 35 sensor streams, and that it received approval for three additional patents for its AI-enabled geothermal design and drilling technology in the last six months. It plans to grow its delivery footprint and expand into the Northeast and Midwest. The company says its solution delivers behind-the-meter thermal energy at up to 50% lower cost, and that a first-of-a-kind Digital Thermal Response Test saves developers nearly $100,000 in the early feasibility and design phase. The funding is not for a model. It is for AI that has to work in the ground. m:…2591
What to watch
There is a caution attached to the enthusiasm. On 20VC, Almeida said he believes AI has been a bubble for a while and that he plans for an eventual bubble burst. That is the discipline this kind of bet needs: a harness company earns its keep only if agents stay unreliable enough to need one, and only if its techniques keep working as models improve. One early signal cuts in Noeri's favour. When Cursor released Composer 2.5 in May, it listed Reinforcement Learning via Self-Distillation, a paper Noeri's founders co-authored, among the work behind the model's new training method. For a venture investor, the decision is now a choice between two theses: fund the next model, or fund the layer that makes the current ones dependable. This week's money picked the layer. m:…6725 m:…3471
What is still unresolved
- The terms and valuation of Noeri's $110 million fundraise were not disclosed in the announcement.
- Whether Noeri's results on the problems it attempted generalise to production agentic work.
- Whether test-time training stays cost-effective as base models improve.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.