AI Agents Hit $2.9B ARR as Trust Gap Emerges in Autonomous Systems

AI agents aren’t a future trend—they’re already pulling in nearly $3 billion in combined annual recurring revenue. But as these autonomous systems make increasingly consequential decisions, the infrastructure to verify their actions hasn’t kept pace.

A detailed analysis from EigenCloud’s blog breaks down the numbers: Cursor crossed $2 billion ARR in February 2026, doubling in three months. Harvey’s legal AI hit $195 million ARR after tripling revenue in 2025. Replit Agent scaled to approximately $150 million. Sierra Agent surpassed $150 million. The list continues with Fin, Cognigy, Devin, and others.

What makes this different from previous enterprise software waves? These aren’t assistants waiting for prompts. Cursor now orchestrates entire coding workflows—reading requirements, writing code, running tests, and shipping to production. Harvey drafts settlement language on $50 million cases. Cognigy processes over one billion customer interactions annually.

The Economics Are Staggering

Anysphere, Cursor’s parent company, generates roughly $13.7 million in ARR per employee. Salesforce generates about $290,000. That’s 47x revenue efficiency compared to the largest enterprise software company on earth.

Intercom’s Fin charges $0.99 per resolution—not per message, per actual resolved ticket. Human agents cost $15-25 per resolution when factoring in salary, benefits, and overhead. That’s a 15-25x cost compression that creates what the EigenCloud analysis calls “an economic inevitability.”

Klarna’s internal AI deployment replaced 700 full-time customer service equivalents, generating approximately $60 million in annual cost savings. The company used this as a key narrative in its IPO process.

Where Trust Breaks Down

Code can be reverted. A bad customer service response gets corrected. But when an autonomous logistics agent reroutes a $2 million pharmaceutical cold-chain shipment and the temperature breaks, people can die. When Harvey’s agent hallucinates a legal precedent in settlement language, a $50 million case collapses.

The AI in logistics market hit $17.96 billion in 2024 with projections reaching $707.75 billion by 2034. By 2026, 62% of executives expect AI agents will make autonomous decisions in supply chain operations, according to IBM research cited in the analysis.

Current verification approaches each have limitations. ZK proofs remain impractical for large neural network inference—proving a single forward pass of a 70 billion parameter model would require orders of magnitude more compute than the inference itself. Optimistic fraud proofs require re-execution, but GPU inference is non-deterministic due to floating-point operation ordering and kernel race conditions.

EigenCompute’s Approach

EigenCloud’s solution uses hardware-isolated execution via Trusted Execution Environments. When you deploy to EigenCompute, your Docker-packaged application loads into an enclave where TDX hardware generates a cryptographic attestation—a signed hash of the exact code running inside. Every deployment records permanently onchain by its Docker digest.

The platform is secured by $17.5 billion in restaked assets. If an operator produces fraudulent attestation, they face automatic economic penalties through slashing.

For the agent categories already generating billions, this creates specific use cases. Legal agents get an evidentiary chain showing which model version and inputs produced a document. CX agents get auditable decision trails. Logistics agents—where no standalone product has broken out like Fin or Devin yet—could finally have the trust layer needed for high-stakes freight decisions.

What Comes Next

The market is past asking whether AI agents will make money. The combined $2.9 billion ARR from named agents with verifiable pricing models answers that question. Average AI agent companies trade at 52x ARR according to AI Funding Tracker data. Total VC investment in AI hit $340 billion in H1 2026 alone per Silicon Valley Bank.

The harder question: as agents get more autonomous and decisions get more consequential, who verifies they did the right thing? Right now, the answer is nobody. That gap between what agents are doing and what anyone can prove they did is where the next infrastructure battle will be fought.

Image source: Shutterstock