ReasonBlocks
ActiveThe runtime layer that makes AI agents cheaper and more reliable
About
ReasonBlocks is the runtime layer that makes AI agents cheaper and more reliable. We correct agent reasoning and compress tokens mid-run, with gains compounding across runs. On SWE-Bench pro we have seen a 52% token reduction and 42% accuracy lift.
Founders · 2
Formerly at Stanford CS. Published in Nature Sustainable Agriculture (ML × agriculture). UN Best Paper. USAMO qualifier. 11 years building with co-founder Rohan. Currently making AI agents smarter and more reliable @ ReasonBlocks
Formerly at CMU, studying Information Systems & Artificial Intelligence. Molecular dynamics + AI research. AI for distributed energy at ENGIE. IoT research at UC Davis. 11 years building with co-founder Sajeev; we're making LLMs cheaper and more reliable.
Launch
Infra layer for smarter and cheaper AI agents
ReasonBlocks provides a runtime that detects production AI agent failures, compresses irrelevant context, and builds a private reasoning library from each run to improve future performance. It targets teams deploying AI agents across domains (legal, finance, healthcare, security, research) and promises cheaper, more reliable runs as usage grows.
Related startups
Reasoning and verification infra for AI
The default way of running on-device AI at Scale
