The frontier-lab mega-rounds get the coverage — the billion-dollar inference bets and the escape-hatch financings. But the two most instructive agent rounds of the summer closed quietly, 48 hours apart in late June, and they funded neither a model nor a framework. Read together, Taktile's $110M Series C and 8090's $135M Series A are the same bet, and it's a bet worth understanding if you're building.

What each one is#

Taktile (June 24, $110M, led by Growth Equity at Goldman Sachs Alternatives) sells an Agentic Decision Platform to banks and insurers. The workflows are the ones where a wrong answer is expensive and audited: business-loan underwriting, claims assessment, anti-money-laundering checks. The company reports 95% automation in B2B underwriting and 75% fewer AML false positives. The product is not a model — it's the governed layer where a regulated institution can let an agent make a call and still answer to an auditor.

8090 (June 26, $135M, led by Salesforce, founded by Chamath Palihapitiya) sells a Software Factory: teams of people and AI agents building and changing enterprise software together, turning user-written documents into code. It explicitly uses third-party agents — model-agnostic by design — and serves healthcare, insurance, aerospace, energy, financial services, and the US government. The product is not an agent — it's the human-in-the-loop orchestration around agents doing real, high-stakes work.

Why they're the same bet#

Neither round funded a horizontal framework (that layer — LangGraph, CrewAI, the vendor SDKs — is mature and already well-capitalized). Neither funded a new model. Both funded a vertical: one specific, high-stakes, regulated workflow, wrapped in governance, audit, and human oversight.

Investors are paying for the trusted, domain-specific layer on top of agents — not the agents.

That's the whole signal. The value isn't accruing to the thing that makes an agent act. It's accruing to the thing that makes an institution trust an agent to act: the audit trail, the oversight surface, the domain-specific guardrails a general framework can't ship for you.

What's left for a founder to build#

The read is unusually clear. The horizontal land grab is over and funded. The open ground is the governed vertical — and it's open precisely because the moat there is domain knowledge and trust, not model training or infrastructure spend.

Pick one regulated, high-stakes workflow. Automate it with off-the-shelf agents and cheap inference. Then spend your real effort on the compliance, audit, and human-oversight layer around it — the surface Taktile and 8090 just raised nine figures to own in their verticals. That surface is fundable, it's defensible, and, because the agent tooling underneath it is now commodity, a two-person team can build it for a fraction of what it would have cost two years ago.