Coding Agents Get Production Funding… and Research Says They Aren't Quite Ready.

Week of September 3 – 9, 2026. Follow the money.

The biggest single check this week went to a model lab. Mistral raised about $3.5 billion to keep building open-weight models companies can run themselves. But the money worth watching went to the companies building agents that run on those models. Cognition raised over $2 billion for the coding agents behind Devin and Windsurf. Harvey raised $550 million for legal work. Both rounds are priced for production. And the same week, a run of new research measured what these agents actually do on real tasks. The valuations and the measurements don’t agree.

Film by Webster AI

Coding agents got valuations that assume they already work

Cognition raised over $2 billion at a $48 billion valuation. Harvey raised $550 million at $15.5 billion. Numbers like that don’t pay for a promising demo. They assume the agents are deployed and trusted on real work at real companies right now. That’s the bet the money is making: coding and legal work handed to agents, in production, at scale.

The infrastructure money went to chips, optics, and power

The week’s other big checks went to the hardware the applications run on. Analog Devices bought Alif Semiconductor for $1.35 billion in cash for its edge-AI chips. Gimlet Labs raised $300 million for software that splits an inference job across different kinds of chips. Celero raised $275 million for the optical parts that carry data between AI data centers. And Bluecore raised a $50 million seed, eight weeks after launch, for floating nuclear reactors aimed at ports, which is really a bet on where compute power comes from next. The smaller applied rounds kept going narrow at the same time: Forus took $150 million for healthcare agents that handle prior authorization, and Clay raised $115 million for a sales and marketing data platform.

The security money followed the agents too…

Cymphony raised $25 million for security that tracks what AI agents can reach and do inside a company. HelmGuard raised $7.3 million for agent governance and risk checks. It’s the same pattern from the last few weeks: the money isn’t paying to make the models safer, it’s paying to control what the agents around them can touch, carry, and spend. The controls are getting funded right as the first real incidents start showing up.

Research

Three papers landed this week, and together they measure the gap between what the agents are funded to do and what they actually do.

Asked to build a working customer-service agent under real conditions, the strongest setup tested — Claude Opus 5 running in Claude Code — passed 23.9% of the evaluations against an expert-built reference. Not most of them. A quarter. The exact job the money is most excited about, agents building and running software, is the one the best agent gets wrong three times out of four.

A second paper checked whether an agent can even tell you how far along it is. Mid-task, the agents reported their own progress correctly between 5.8% and 11.5% of the time. They’re reliable at the very start and the very end and mostly wrong in the middle, which is the stretch where you’d actually want to trust the status.

The third looked at the benchmark numbers vendors quote. On a cleaned-up version of SWE-Bench Pro that blocks the evaluation shortcuts, one leading model dropped from 78.8% to 57.3% once it couldn’t game the test. A lot of the headline coding-agent scores are measuring the benchmark, not the agent.

The takeaway: We know that even research trails the money when it comes to the frontier. I think the opportunity here is to be wary and cautious when trusting agents to do real work. A framework should be in place, along with evals and governance, to ensure that when they do fail (and they probably will), somebody is there to catch them and make improvements.

Sources

Every funding fact above is a verified round from a primary source, confirmed for the week of September 3 – 9, 2026.