AI Agents Get Formal Verification, Digital IDs, and a Rebrand as Startups Race for Reliability

As enterprises push AI agents from pilots into production, reliability has become the decisive factor. Three separate announcements this week illustrate the depth of the challenge: startups are raising tens of millions to make agents trustworthy, governments are planning digital identities for them, and researchers are teaching robots to train themselves through AI coding agents.

On Wednesday, Pramaana Labs announced a $27 million seed round led by Khosla Ventures to bring formal verification to AI. The startup combines large language models with a deterministic verification layer built on the open-source LEAN programming language, targeting verticals like tax preparation and drug discovery where hallucinations cannot be tolerated.

Formal verification meets AI agents

Pramaana co-founder and CEO Ranjan Rajagopalan frames the problem as one of codification. The company is working with former IRS commissioner Danny Werfel on tax law use cases and with professors from IIT Delhi, IIT Madras, and UC Berkeley on cybersecurity and drug discovery. For each vertical, domain experts build a LEAN-style formal system that overrides the LLM's probabilistic output when it violates known rules.

• Seed round: $27 million from Khosla Ventures, Accel, BoldCap, Nexus Venture Partners, Premji Invest, and Unbound.
• Method: LLM engine plus deterministic formal verification layer using LEAN language.
• Target verticals: law, drug discovery, tax preparation. Reference project: France's CATALA for tax code formalization.

Redefining the agent as a teammate

Andreessen Horowitz is betting $38 million that the industry needs a new vocabulary. That money went to Convey, a startup that deliberately avoids the word "agent" and instead pitches AI "teammates." The Series A round also included Khosla Ventures and Pear VC. Convey argues that agents currently impersonate users during tasks, creating security and identity problems. Its alternative is a persistent, permissioned AI entity that operates alongside the user.

Digital identity for AI agents

Estonia plans to solve the impersonation problem more directly. Prime Minister Kristen Michal announced the country will issue personal identification numbers to AI assistants, making it the first nation to do so. When an agent acts on behalf of a user, it will use its own ID rather than borrowing the human's credentials. The move builds on Estonia's existing digital ID infrastructure and could become a template for other nations.

Robots that train themselves

Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley demonstrated last month that AI coding agents can teach robots dexterous grasping in the real world. A fleet of eight robots achieved up to 99% success on challenging manipulation tasks by autonomously generating and executing training code. The approach reduces the need for human-labeled data and could accelerate commercial robot deployment.

Together, these developments point to a maturing ecosystem: formal guarantees for high-stakes decisions, identity standards for agent interactions, and self-improving physical systems. Connectivity remains an open question. AST SpaceMobile's launch of three more Bluebird satellites this month suggests that direct-to-device links could eventually keep agents online anywhere, but the foundational work is happening in code and regulation.