Thariq Shihipar describes advanced models as uneven systems whose abilities must be discovered through practice. He uses examples where a model performs poorly in ordinary chat but succeeds once a coding harnessAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction. gives it tools, data and room to inspect the resultContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task., a gap he calls capability overhang.
Before implementation, Shihipar recommends a blind-spot pass that searches the codebase and asks which unknowns could change the planProblem framing with AI is the human-led process of defining the real goal, constraints, evidence and success criteria before using a model to explore solutions.. Interviews, quick prototypes and concrete visual references help convert preferences that are difficult to describe into evidence the model can use. During execution, implementation notes should record deviations and unresolved choicesAn agent audit trail is an ordered, durable record of an AI agent's instructions, decisions, tool calls, state changes, identities, and outcomes. so the human stays informed.
Shihipar acknowledges that faster agentic programming can feel like losing a valued craft, but argues that builders should respond by increasing ambition and learning the new workflow. Cheap code is not the goal. The test is whether repeated experiments produce real value and leave people more capable. Event promotions are omitted.
Watch on YouTube
