Context engineering with Dex Horthy

Gergely Orosz interviews Dex Horthy about context engineering, coding-agent loops, and software factories ([00:01:00]-[00:02:03], [00:20:01]-[00:21:04], [01:20:59]-[01:22:02]). Horthy argues for controlling model inputs, bounded loops, and human judgment before consequential implementation ([00:21:04]-[00:23:00], [00:34:35]-[00:37:43], [00:58:23]-[01:00:30]).

Key Points Covered

• Context engineering controls the model's actual input: Horthy looks beneath abstractions such as retrieval, memory, agent history, and structured output. He controls token selection and ordering. He presents this as a route toward higher reliability, not a guarantee ([00:20:01]-[00:23:00]).
• Information and instructions have separate budgets: Long contexts can contain irrelevant information, conflicting instructions, and abandoned conversational paths that make the next action less dependable ([00:28:06]-[00:31:03]).
• The outer harness includes the whole working environment: Horthy uses the term for commands, MCP tools, skills, repository organization, and development setup around an existing agent. He distinguishes this environment from the inner harness's tool interfaces ([00:26:00]-[00:27:02]).
• Loops need strong back pressure: Horthy says compilers, tests, linters, or measurable performance should constrain progress. His team's reported “slow loop” fixes a few issues. It opens reviewable pull requests rather than producing one large autonomous change ([00:33:32]-[00:37:43]).
• Passing tests can hide maintainability failures: In HumanLayer's reported 2025 experiment, unattended feature generation created a cross-cutting identity-type problem. Tests and agent review did not expose it. Humans eventually had to recover the codebase context and diagnose it ([00:41:23]-[00:46:11]).
• Planning artifacts should be disposable context: Horthy says research and design can compact codebase state and intent, but detailed plans duplicate review effort and persistent specifications drift from code ([01:01:32]-[01:06:43]).
• Compact deliberately between difficult phases: His current workflow starts fresh contexts between research, design, and planning, concentrating human involvement on architecture and vertically testable implementation slices ([01:06:43]-[01:09:36]).
• Automate bounded subloops, not the whole factory: Orosz frames implementation and review as stages in an existing product-delivery loop. Horthy argues for adding checkpoints instead of removing human review from the system ([00:49:00]-[00:52:08], [00:58:23]-[01:00:30]).

Full video: https://www.youtube.com/watch?v=Usufn8IQJgw(opens in a new tab)