Dex Horthy

Dex Horthy is the co-founder and chief executive of HumanLayer, which builds collaborative AI-powered development tools for engineers working in complex codebases. He created 12-Factor Agents and helped establish context engineering as a practical discipline for making coding agents reliable without sacrificing architectural judgment or maintainability.

From enterprise infrastructure to coding agents

Horthy started programming in high school at NASA’s Jet Propulsion Laboratory. He subsequently spent seven years at Replicated, progressing through engineering, solutions engineering, product, and executive responsibilities. His work included container orchestration and helping companies including HashiCorp, DataStax, and H2O.ai deliver self-hosted Kubernetes products.

HumanLayer joined Y Combinator’s Fall 2024 batch. Its initial product let agents request human approval or clarification through Slack, email, and SMS—a response to Horthy’s experience building automation capable of taking consequential actions, such as deleting unused database tables. Initially a solo founder, he later brought on Kyle Mistele as co-founder and chief technology officer. The company expanded from approval infrastructure into a collaborative engineering environment for teams adopting coding agents.

• 12-Factor Agents makes reliability an engineering problem. Horthy’s open-source framework treats agents as conventional software augmented by carefully selected model decisions. Engineers should control prompts, context, execution state, and orchestration; interpret tool calls as structured outputs executed by deterministic code; and build workflows that can pause for human approval. Small, focused agent loops belong inside dependable systems, not above them.

• Context engineering depends on evidence, not volume. Horthy argues that incorrect information, missing details, stale documentation, and noisy tool output degrade agent performance. His practice of intentional context compaction distills repository research into concise, reviewable artifacts containing the relevant files, behavior, and implementation constraints. Subagents help by investigating independently and returning compact findings; assigning them fictional organizational roles adds little.

• Research, Plan, Implement preserves shared understanding. His workflow for established codebases separates investigating the actual system, preparing explicit implementation and verification steps, and executing the reviewed plan. Beyond improving model output, the approach maintains human mental alignment: engineers can understand how their codebase is changing before reviewing extensive generated diffs. Horthy demonstrated the method on the BAML Rust codebase while acknowledging its limits when an attempt to remove Hadoop dependencies from Parquet-Java required renewed architectural analysis.

• Passing tests cannot guarantee long-term maintainability. After HumanLayer experimented with highly autonomous development, Horthy sharpened his critique of lights-off software factories. Coding benchmarks and reinforcement-learning rewards can favor immediate task completion while overlooking coupling, poor program design, and architectural damage that surfaces months later. His alternative combines product review, architecture, explicit program design, vertical implementation slices, and continued human code review. He has also argued publicly that automated linting and AI review cannot yet replace reading code.

Horthy continues experimenting directly with developer tooling, including shannon, which manages interactive coding-agent sessions through tmux. His central conviction is that faster software delivery remains valuable only when engineering teams retain responsibility for the systems they ship.