▶ Watch ↗Popular #520:31
Conference affiliation: Co-Founder · HumanLayer · 2026
Dex’s September 18 post sharing the recording and emphasizing code review and software-factory design.
x.comDex Horthy is a member of the technical staff at Factory and the founder of HumanLayer. He started HumanLayer so software agents could ask permission before acting, then helped turn it into a workspace where engineering teams could review what coding agents understood, planned, and changed. He also created 12-Factor Agents, a guide to putting model decisions inside software whose execution engineers control. On November 4, 2026, Factory announced its acquisition of HumanLayer. Dex and co-founder Kyle Mistele joined Factory to bring HumanLayer’s collaboration features to its platform. Dex’s public profile · Dex and Kyle’s announcement · Factory’s announcement
During a 2008 summer internship at NASA’s Jet Propulsion Laboratory, Dex built a graphical tool for finding routes across the Moon’s heavily cratered south pole. The tool accounted for how steep a slope a rover could climb. At Sprout Social from 2013 to 2016, he maintained internal APIs, automated infrastructure provisioning and deployment, added monitoring, and led refactors of legacy Java services. He joined Aspiration as a backend engineer and went on to lead platform engineering. Career history
Over nearly seven years at Replicated, Dex moved from building deployment infrastructure to helping vendors ship software that customers could run in their own environments. He started the team that prototyped and shipped an embedded Kubernetes installer that preceded kURL. As Replicated’s first solutions engineer, he worked with CEO Grant Miller to pitch and onboard customers, then led solutions engineering and customer success. In product management, he worked with the Vendor Experience team to make customer-hosted applications easier to monitor and advised the CEO. He also advised Prefect. Dex’s account of his time at Replicated · Career history
HumanLayer grew out of Metalytics, a data engineering venture Dex started with a collaborator. Their recommendation prototype could flag warehouse tables that had gone unqueried for more than 90 days. The query logs showed when a table was last used. Deciding whether to keep it still fell to a person. Dex and his collaborators added a step that asked for approval before the agent ran destructive SQL. Dex then started HumanLayer, which turned that approval step into infrastructure other developers could use to request permission, clarification, or feedback through Slack, email, or SMS. Metalytics and HumanLayer interview · HumanLayer at Y Combinator
Dex entered Y Combinator’s Fall 2024 batch as HumanLayer’s solo founder and CEO. Kyle joined as co-founder and CTO in 2025 after the two had traded ideas and worked together at a hackathon. Separately, at WeaveHacks, Dex watched Kyle independently build an OAuth prototype for Model Context Protocol connectors across customer accounts. By the time they seriously discussed working together, Kyle was already using the CodeLayer alpha daily. Dex’s announcement of Kyle as co-founder
One of Dex’s early DevOps agents kept running Makefile commands out of order. After two hours of revising the prompt, he had spelled out a sequence a short shell script could have run. That experiment and conversations with other builders informed 12-Factor Agents, which Dex created for developers introducing model-driven decisions into existing software. Its patterns give engineers control over what goes into a model request and how the application handles the response. The DevOps experiment · 12-Factor Agents
A wrong description of a codebase can send an entire implementation in the wrong direction. Dex’s practice of intentional context compaction preserves the relevant files, behavior, and constraints in a short document a person can check and a fresh agent session can use. Subagents can do the searching in separate sessions and return focused findings. Investigating the code afresh takes time, but it gives engineers an account they can check against the code they are about to change, even when onboarding documents have gone stale. Context engineering for complex codebases
Dex and the HumanLayer team built Research, Plan, Implement around that handoff. Research records how the system works today. The plan sets out the changes and how to verify them, giving teammates a chance to question the design before the agent writes code. Dex calls the shared understanding this creates human mental alignment. More detail can help the agent execute a plan and make it harder for a teammate to read. A button-color change can go straight to the agent. A change across several repositories can justify a plan the team reviews together. Research, Plan, Implement in practice
Dex and Vaibhav Gupta, his co-host on AI That Works, compared plans made with and without prior research on BAML’s Rust codebase, then used the approach on larger changes together. An attempt to remove Hadoop dependencies from Parquet-Java exposed a limit. Once Dex and his collaborator understood the architectural problems, they discarded the research and plans and went back to the whiteboard to decide how the pieces should fit. The BAML and Parquet-Java experiments
In July 2025, HumanLayer tried letting agents write code without anyone reading it. When automation could no longer solve a problem, Dex had to investigate a codebase he had stopped keeping up with. A test can show that a patch fixes a bug today while missing a design choice that makes two parts of the system harder to change independently. The cost may surface months later, when a change in one part breaks the other. Dex argues that this delay makes maintainability hard to capture in model training feedback. He also notes how hard it is to measure improvement without benchmarks that test what happens as the code changes. Why Software Factories Fail
For his October 15, 2026 report, Dex ran coding models on SlopCodeBench, developed by Gabriel Orlanski and colleagues. The benchmark introduces requirements in stages, testing whether a model can add new behavior while preserving what worked before. Dex wrote up the comparison and credited Gabriel’s contributions to it. SlopCodeBench paper · Dex’s benchmark report
For substantial changes, Dex favors agreeing on product behavior and architecture early, then reviewing program design down to types, method signatures, and how components call one another. Implementing in vertical slices gives reviewers working behavior to check along the way. He still favors bounded automation, such as sending new pull request comments to an agent for follow-up. In an August 2026 response to Robert C. Martin, Dex said he remained skeptical that deterministic linting and AI review could replace engineers reading code. Why Software Factories Fail · The Great Loops Debate · Dex’s response to Martin
HumanLayer’s later product put agent sessions, research, plans, worktrees, and diffs in a shared workspace. A teammate could comment on a design while it was still being written and send that feedback to the agent, or inspect a live diff before a pull request opened. On October 23, 2026, Dex and Kyle released HumanLayer’s core under Apache 2.0, including the desktop and web apps, session runner, and sync layer. Dex also built Shannon, a smaller tool that runs interactive Claude Code sessions in tmux and exposes their output through a CLI and SDK. HumanLayer workspace · Open-source release · Shannon
At Factory, Dex and Kyle plan to bring shared tasks and workstreams, comments on plans and design documents, and live code changes to the platform. At the time of the acquisition announcement, the founders committed to keeping the hosted HumanLayer service online for at least 30 days and offered users help exporting data, moving to Factory, or self-hosting. Dex and Kyle’s announcement · Factory’s acquisition announcement
Topics covering these talks
▶ Watch ↗Popular #520:31
▶ Watch ↗Popular #1317:06
▶ Watch ↗AI Engineer World's Fair 20261:00:16
Loading more talks and highlights…