The Pragmatic Engineer — selected conversations
Context engineering with Dex Horthy
Eighteen curated conversations on agent workflows, design, harnesses, and software quality
A manually reviewed selection of Pragmatic Engineer conversations with Maggie Appleton, Matt Pocock, Tibo Sottiaux, Martin Fowler, Dex Horthy, Boris Cherny, Mario Zechner and Armin Ronacher, Laura Tacho, Hillel Wayne, Mitchell Hashimoto, Kent Beck, Peter Steinberger, Grady Booch, Steve Yegge, DHH, Dax Raad, Charity Majors, and Addy Osmani. Topics include design prototyping, reusable skills, agent workflows, context and harness design, TDD and verification, architecture, measurement, accountability, and open-source governance. Practitioner observations, product-team claims, and forecasts are not controlled findings.
In this collection Browse 18 summaries 5 of 18
Gergely Orosz interviews Dex Horthy about context engineering, coding-agent loops, and software factories 01:00–02:03(opens in a new tab), 20:01–21:04(opens in a new tab), 1:20:59–1:22:02(opens in a new tab). Horthy argues for controlling model inputs, bounded loops, and human judgment before consequential implementation 21:04–23:00(opens in a new tab), 34:35–37:43(opens in a new tab), 58:23–1:00:30(opens in a new tab).
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 20:01–23:00(opens in a new tab).
- 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 28:06–31:03(opens in a new tab).
- 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 26:00–27:02(opens in a new tab).
- 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 33:32–37:43(opens in a new tab).
- 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 41:23–46:11(opens in a new tab).
- 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 1:01:32–1:06:43(opens in a new tab).
- 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 1:06:43–1:09:36(opens in a new tab).
- 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 49:00–52:08(opens in a new tab), 58:23–1:00:30(opens in a new tab).
Full video: https://www.youtube.com/watch?v=Usufn8IQJgw(opens in a new tab)