Raymond Weitekamp is the founder and chief executive of polySpectra, a materials scientist and manufacturing entrepreneur developing recursive coding agents through independent research and OpenProse. His work addresses a problem familiar from both advanced manufacturing and autonomous software development: converting impressive underlying capabilities into reliable, repeatable production systems.
Weitekamp studied chemistry at Princeton and earned a doctorate in chemistry at Caltech, where he worked with Robert Grubbs and Harry Atwater on photonic materials and polymer chemistry. He developed a light-activated ruthenium catalyst associated with Grubbs’s work in olefin metathesis, then joined the founding cohort of Cyclotron Road at Lawrence Berkeley National Laboratory.
At the laboratory’s Molecular Foundry, he tested whether his chemistry could produce useful 3D-printed parts before pursuing investment. That work became polySpectra, an independent company making rugged photopolymer resins for additive manufacturing. He also co-founded Cypris Materials. His distinction between scientific risk—whether something works—and engineering risk—whether it works reliably, repeatedly, and affordably—also frames his approach to coding agents.
Making autonomous coding agents dependable
Through RAW.works, Weitekamp progressed from AI-assisted reading and software workflows to agent automation, recursive reasoning, memory, and orchestration. At AI Engineer World’s Fair 2026, he represented OpenProse and outlined his approach to recursive coding agents: agent failures often reflect inadequate management, delegation, tooling, and verification more than insufficient model intelligence.
His research concentrates on four connected ideas:
Recursive Language Models as executable reasoning systems. Large inputs remain outside the immediate context window while a model inspects, partitions, and transforms them through an executable environment. Crucially, the model determines how to decompose the problem; fixed MapReduce pipelines lack that adaptive behavior.
Agent memory through recursive scaffolding. His DSPy and LongMemEval experiments investigated recursive harnesses as memory systems, finding that structured outputs, typed observations, delegation, and model selection materially influenced results. His LongCoT experiments also explored whether smaller models could tackle demanding reasoning tasks through recursive execution.
Self-delegation with operational guardrails. His open-source project ypi enables Pi coding agents to invoke equivalent agents recursively, with controls covering depth, spending, runtime, and calls. Complementary tools include rlm-cli for command-line recursive-model workflows and dirpack for compact, budget-constrained directory representations.
Explicit skills and reusable execution contracts. At OpenProse, he added explicit skill and tool dependencies for delegated agents and built workflows that convert successful coding sessions into repeatable execution plans. His Auto-Pocock workflow applies those ideas to engineering tasks that move from a feature brief through verification.
Reliable coding agents need more than intelligence: they need ways to decompose work, preserve state, verify delegated results, and reuse successful workflows.
