Raymond Feng is an Applied Compute researcher and software engineer developing AI agents that improve through experience inside real enterprise workflows. His work spans continual post-training, reinforcement-learning infrastructure and organizational memory, with an emphasis on adapting models to the environments where they actually operate.
Feng studied computer science at the University of California, Berkeley, graduating in 2020. He built financial-data pipelines and internal trading tools at World Tree Technologies, developed cloud-connected security software at Turing Video, and worked on enterprise security, onboarding automation and billing infrastructure at Google Cloud. His professional background and guide to deploying multicontainer web applications reflect practical experience building production systems before his move into agent research.
How agents learn from real work
- Task-level training: Feng treats complete tasks as the meaningful unit for improving agents, connecting final outcomes to model-generated decisions across tool calls, multiple conversations, cooperating subagents and context compaction. His work on multiagent reinforcement-learning infrastructure addresses workflows that cannot be evaluated adequately one response at a time.
- Environment fidelity: Training infrastructure creates behavioral incentives even when developers never intended them. Failed tool calls can encourage artificially short responses, while discarded timeout trajectories can teach agents to stall difficult assignments. Feng argues for training inside existing enterprise agent harnesses, while acknowledging that real customer interactions usually cannot be replayed and often yield qualitative feedback instead of clean numerical rewards.
- Relevance-Masked Self-Distillation: Feng and Pranav Vaid developed targeted self-distillation to teach unfamiliar, company-specific behaviors while preserving existing capabilities. Their approach emphasizes the model-generated tokens most relevant to the desired behavior, providing a useful training signal when conventional reinforcement learning produces too few successful examples.
- ACL-Wiki: Feng coauthored research on agent memory grounded in production traces, describing a system that transforms coding-agent interactions into reusable organizational knowledge. Human feedback helps refine memories that later agents can draw on during software-development tasks.
His longer-term goal is agents that evaluate their own interactions and update their capabilities continuously, reducing the need for engineers to identify and repair individual failures manually.
