Musings on engineering, cybersecurity and AI. Kyle Mistele RAG is (Not) Dead: How to Think about Building RAG Systems Building Kydenticon: A TypeScript Library for GitHub-Style Identicons You're Doing it Wrong: Prompt- and Context-Engineer with XML, not JSON MCP Deep Dive: the Great, the Broken, and the Downright Dangerous Why Everybody is so Excited about MCP
Kyle Mistele is co-founder and chief technology officer of HumanLayer, where he develops infrastructure for using coding agents on complex production systems without sacrificing architectural judgment, code review, or operational control. Previously an offensive-security engineer and CTO of Naptha AI, he applies experience in cybersecurity, distributed systems, and developer tooling to a central question: how much work can agents perform while software remains understandable and safe to change?
From offensive security to agent infrastructure
Mistele studied computer science and cybersecurity at Southern Methodist University and earned the Offensive Security Certified Professional certification in 2020. His security career included red-team engineering and ATM hacking; personal projects include bladerunner, a collection of evasive shellcode runners, and wifi-squid, which handles protected wireless credentials.
He subsequently worked on multi-agent systems as CTO of Naptha AI and built a model-agnostic TypeScript adaptation of OpenAI’s Swarm framework using the Vercel AI SDK. At HumanLayer, he joined founder Dexter Horthy as CTO and co-founder after their work intersected around the Model Context Protocol, developer events, and coding-agent infrastructure.
Technical positions and defining projects
Agentic control loops for production code. Mistele frames automated software changes as a measurable control system: identify deviations from a desired codebase state, select a bounded intervention, implement it, and evaluate the result. His approach to loop engineering favors small, reviewable pull requests over unlimited code generation. HumanLayer applies this approach to an incremental RPC migration to Effect, using ast-grep to identify unmigrated procedures, version-controlled baselines to prevent regressions, handwritten examples to guide agents, and production telemetry to prioritize improvements. Scheduled workflows perform one iteration at a time, retain reviewer feedback, and pause when earlier pull requests remain open.
Deterministic tools before autonomous improvisation. Repository scans, sorting, scheduling, and pull-request creation should remain ordinary deterministic software when possible; agents should handle tasks requiring interpretation or implementation. The goal is to control token costs, review burden, and maintenance risk without discarding useful automation.
Retrieval is an information-retrieval problem. Mistele rejects the assumption that retrieval means vector search, recommending lexical search, metadata filters, syntax-aware code search, or hybrid pipelines when they better match actual documents and queries.
MCP security and authorization boundaries. His analysis of Model Context Protocol implementations examines transport limitations, OAuth architecture, executable dependencies, and the privileges granted to local and remote integrations.
His smaller projects include Kydenticon, a zero-dependency TypeScript library for deterministic GitHub-style avatars. In a public assessment of AI-assisted systems design, Mistele draws a sharper limit: models can generate options faster than people without reliably producing strong distributed-systems abstractions. Human judgment remains responsible for the architecture.
