We've opened Pandora's box on AI coding agents and the instinct is to react with hype or fear. This talk offers a third option: read the instructions in the box.

Agent harnesses moved from research demos to production tools in twelve months. The system is far less mysterious than the discourse suggests: a single while-loop calls a model and dispatches tools, with deliberate engineering all around it. Roughly 1.6% of the codebase is AI decision logic. 98.4% is operational infrastructure.

But at least six 'ills' did fly out: opacity (your diff captures the *what*, not the *why*), misalignment (OpenAI runs internal monitors at maximum reasoning to flag escalations), displacement (Anthropic reports 27% of agent-assisted tasks wouldn't exist without the tool; both ambition and baseline rise), accountability vacuum (who's accountable when the agent decided?), amplified dysfunction (AI multiplies whatever system it runs on, including its sins), and compliance-as-bolt-on (you'll discover your governance gaps in production, not in design).

Inside the box, safety primitives are already shipping (permissions, hooks, MCP boundaries, subagent isolation, append-only sessions), and the engineers who compound are the ones who learn to read them. The only durable strategy is agency: shape your harness, don't just prompt it. When AI flattens skill differences, the developers who shape get disproportionate leverage.

The audience will leave with five concrete moves to make Monday morning, the literacy to read any production agent harness on its own terms, and a sketch of where the practice scales: the AI-Native Software Factory.

Audience: developers at any level, juniors through staff. Mixed-experience room.