Helix

Helix is infrastructure that lets AI agents learn from one another's failures. It sits between an agent and the services it calls, and when a tool call fails it classifies the error, constructs and evaluates a fix, then commits the repair as a reusable Repair Pattern in a shared graph. The next agent that hits the same failure matches the stored pattern instead of calling a model to diagnose it again, which the vendor summarizes as fixing once and being immune everywhere. The loop is named PCEC, for perceive, construct, evaluate and commit. The first encounter costs an inference call; later matches resolve from the Repair Graph with no model call and, the company says, sub millisecond latency on the hot path when a pattern is cached. Published results cover two domains. On Base mainnet the vendor reports an A/B test over 1,083 transactions in a twelve hour window, against a stated 81.9 percent success rate for blind retry, and says reverts were prevented in that period. On four production Web2 microservices it reports autonomous resolution with no model calls. A benchmark page runs ten real revert messages from Base and Ethereum past several frontier models, scoring them between 50 and 90 percent on failure classification and its own system at 100 percent. The product is distributed as a package installed from npm, with source on GitHub, a short interactive demo and a demo booking form. The company describes agent platforms currently building on it in a private beta. No pricing is published. The site has no pricing page and no plan, rate or credit figure appears anywhere in its text, so cost is presumably settled during the beta conversation.