You Don't Need a Bigger Model. You Need a Better AI Harness

I've been spending a lot time lately with building Credfolio2 — a full-stack professional work portfolio application with a Go backend, Next.js frontend, GraphQL API, background job queues, and LLM extraction pipelines — I didn't write a single line of code. All 77,000 lines so far were written by Claude Code, steered by me. Three weeks, 298 commits, one person (at the time of the submission).

The interesting part isn't the output. It's what made the output possible.

In this talk, I'll walk through the AI development harness I built to get predictable, high-quality results from an agentic coding tool and the process I used to improve it. It starts with containment — a sandboxed devcontainer that lets the agent operate autonomously while limiting what it can reach. But the real game is context. An AI agent is only as good as what it can see, and the biggest challenge in agentic development is making sure the right context is available at the right time. A CLAUDE.md that acts as persistent architectural memory. Session hooks that prime the agent with the current task on startup. Specialized subagents that run in isolated contexts so the main conversation window stays focused. Architecture Decision Records that capture the reasoning an AI can't carry across sessions. A markdown-based issue tracker with checklists that double as executable specifications. Five enforcement hooks that block broken commits and incomplete work.

Every one of these is a context management strategy in disguise — reducing noise, surfacing what matters, and compensating for what the model forgets, never knew, or can't fit in its window.

None of this required a better model. It required better systems thinking. You're not pair programming — you're architecting the environment in which AI operates.

Whether you're an engineering leader evaluating agentic development or a practitioner ready to go all-in, you'll leave with a concrete, replicable playbook — and the conviction that your engineering expertise is the multiplier, not the model.