One researcher used AI agents to design and verify a RISC-V chip in five weeks with zero human-written code

Researchers have demonstrated a novel workflow in which a single person can orchestrate multiple AI agents to autonomously design and verify a working processor chip in just five weeks, with no human-written RTL code and no unverified proofs reaching the final design.

The Salt Method: Verification at AI Speed#

Traditionally, formal machine verification has been an expensive luxury reserved for mission-critical systems because of the cost and time involved. The research shows that when AI agents perform the work, the economics flip—verification becomes not just affordable but essential for maintaining quality and productivity as autonomous systems operate at scale.

The approach, dubbed the Salt method, treats mathematical proof as a currency between AI agents. Instead of allowing any hallucinated or suspect proof to propagate, all mathematical claims are checked against a proof kernel—an incorruptible reference implementation that guarantees correctness. Human attention is focused narrowly on high-level statements, design decisions, and key rulings, rather than on line-by-line code review or proof validation.

From Application to Silicon#

Over approximately five weeks, one researcher working with consumer-grade AI subscriptions directed a fleet of autonomous AI agents through the complete hardware design pipeline:

• Application-level code definition
• Compiler implementation and verification
• Executive layer design
• RISC-V processor architecture
• Fabrication on a community silicon shuttle

No step in this chain required human RTL authorship or manual proof review, yet every stage was mathematically verified and traceable.

Verification Chain and Error Tracking#

The verification trail runs from the Lean 4 kernel (a formal proof assistant) through to SAT-checked equivalence checking at the silicon boundary. The researchers published a complete accounting including theorem provenance, a pre-registered token meter tracking computational work, bounded human-time records, and a detailed error ledger. Over the campaign period (July 7–20, 2026), 256 errors were flagged and caught—with zero incorrect proofs ever reaching the final record. One error number (#79) was never assigned due to implementation details, and later catches went unnumbered, but the monotone counter demonstrates comprehensive error tracking.