Dex Horthy

Dex Horthy explains how intentional context compaction, targeted research, explicit implementation plans, and human review can help coding agents work more effectively in existing codebases without sacrificing team understanding.

• Optimize an agent’s context for correctness, completeness, manageable size, and constructive trajectory; incorrect information is more damaging than missing information or excess noise. 4:52 ↗
• Use intentional compaction to preserve relevant findings, files, and implementation details while starting fresh context windows without repeating exploratory work. 2:52 ↗ 3:59 ↗
• Treat subagents as isolated research contexts that return concise findings, not as fictional frontend, backend, or QA teammates. 6:34 ↗
• Ground research in the current code, then create explicit plans with concrete changes and tests; review those plans to maintain mental alignment and catch mistakes before they multiply. 13:57 ↗ 15:03 ↗ 15:46 ↗ 16:42 ↗
• Scale the workflow to the task: simple changes may need no formal research, while complex or cross-repository work can benefit from deeper investigation and repeated compaction. 17:40 ↗
• Keep humans responsible for architectural reasoning and organizational change: coding agents can amplify sound judgment, but they cannot substitute for it. 10:00 ↗ 16:42 ↗ 19:24 ↗

Dex Horthy explains why dependable agents emerge from structured model outputs, explicit control flow, carefully engineered context, durable application state, focused workflows, and meaningful human participation.

• Use an agent only when model-driven judgment adds value; fully specified, predictable tasks may be better served by ordinary deterministic code. 1:11 ↗
• Treat tool execution as structured JSON plus deterministic application code, and retain direct ownership of branching, looping, termination, and recovery. 4:09 ↗ 5:00 ↗ 6:52 ↗
• Improve reliability through context engineering: inspect prompts, control how state and history are represented, limit irrelevant context, and summarize or remove obsolete errors. 5:53 ↗ 8:36 ↗ 9:29 ↗ 10:27 ↗
• Keep model execution effectively stateless by storing application-owned execution and business state externally, enabling long-running workflows to pause and resume safely. 6:52 ↗ 7:43 ↗ 13:51 ↗
• Embed small, focused agents inside mostly deterministic workflows, and make human approvals or corrections available through channels people already use. 11:19 ↗ 12:07 ↗ 12:59 ↗
• Choose tooling that provides inspectable, owned scaffolding while preserving the flexibility to refine prompts, context, control flow, and human collaboration. 14:38 ↗ 15:16 ↗