← All speakers

Kyle Mistele

Conference affiliation: CTO · HumanLayer · 2026

Kyle Mistele co-founded HumanLayer with Dex Horthy and served as its CTO, helping teams use coding agents to make changes to production software in small, reviewable patches. Both founders joined Factory when it acquired HumanLayer on November 4, 2026. At Factory, the team plans to build shared workstreams and tools for collaborating on plans and documents inside the app.

On this page

Kyle studied computer science and cybersecurity at Southern Methodist University and earned the Offensive Security Certified Professional certification in 2020. He worked in offensive security, including red-team engineering and ATM security testing, before becoming CTO of Naptha AI, where he worked on multi-agent systems. His TypeScript adaptation of OpenAI’s Swarm used the Vercel AI SDK to let developers build workflows that hand tasks between agents while supporting multiple model providers.

Coding agents that work at the pace of review

At HumanLayer, Kyle applied control theory to a problem familiar to anyone who has reviewed a sprawling pull request. A team defines what it wants to change about a codebase, measures where the code falls short, makes a bounded change, and measures again. When HumanLayer began migrating its remote procedure calls to Effect, the team used ast-grep to find procedures still written in the old style. A baseline checked into version control let the team catch new additions to the backlog while a coding agent migrated existing procedures using examples the team had written by hand. Kyle proposed using production telemetry to move procedures with frequent errors or poor instrumentation to the front of the queue. Loop Engineering from first principles

The team switched off an early version because correcting an agent’s patch meant checking out its branch and editing its instructions. In the revised workflow, a reviewer could request another pass from the pull request, and the agent could save the feedback in a version-controlled file for future runs. When unattended patches began to pile up and conflict, the team made the workflow wait for its previous pull request to close before opening another. That decision let the reviewers’ capacity set the pace of the migration. Loop Engineering from first principles

How assistants find information and get access

  • Match the search method to the query. An exact identifier and a question about a concept call for different kinds of search. Kyle argues for choosing among lexical search, metadata filters, syntax-aware code search, and vector similarity based on what users need to find and how the material is structured. A system can combine those methods when queries call for more than one. Kyle’s essay on retrieval
  • Check whose permissions a tool uses. A local Model Context Protocol integration can run code with the user’s permissions. For a remote integration, the OAuth design shapes what the assistant can access on the user’s behalf. Kyle examines how those differences, along with transport and executable dependencies, affect the access users grant when they connect a tool. Kyle’s analysis of MCP

Kyle also built Kydenticon, a zero-dependency TypeScript library for GitHub-style avatars. The same input produces the same avatar, so an application can regenerate it whenever it needs it.

Read the topics behind these talks

1 conference talk

Loading more talks and highlights…

References