It is not tied to one model or harness. Codex, Claude Code, or another capable agent can plug into the same operating system: trace the real codebase, turn direction into a plan, implement against it, prove the result, and carry useful context forward. I keep the product judgment and the final call.
The operating model
One system, three kinds of ownership
The point is not to hand the keyboard to a model. It is to combine human direction, a plan we can inspect, and an agent that can execute with continuity.
I lead
Vision
I define the outcome, the user value, the constraints, and what must not be compromised. The active agent can challenge the direction; it does not invent the mission.
We shape
Plan
The agent traces the current code and retrieves relevant history. Together we turn the intent into files, boundaries, risks, ordered steps, and a clear definition of done.
The agent executes, I steer
Implementation
The chosen harness makes the scoped changes and runs the proof loop. I review decisions, redirect drift, test the experience, and decide what is ready to ship.
The continuity layer
Working memory
The model and harness are replaceable. The accumulated context is the asset. My system keeps useful history local, curated, and source-linked so every compatible agent starts with context instead of amnesia.
MEMORY.md
Index
A searchable map of projects, conventions, decisions, and useful prior work.
rollout_summaries/
Evidence
Compact records of what changed, what was proven, and which boundaries remain open.
skills/
Playbooks
Repeatable workflows for problems that should be solved once and reused, not rediscovered.
The loop
How the AI system works
I own why and what good looks like. The chosen agent helps trace, plan, build, and prove. Working memory makes the next loop—and any compatible harness—smarter.
Start with the why
I bring the product direction: the outcome, the user, the constraints, and the trade-offs I am willing to make. That judgment stays human.
The stack
What the system is made of
The model matters less than the operating discipline around it: real context, explicit ownership, scoped execution, and honest proof.
Agent harnesses
The execution layer is interchangeable. Any capable harness can work from the repository, terminal, browser, and the same plan—without owning the system around it.
Working memory
A local, curated layer of project context, rollout evidence, decisions, and playbooks. It lets useful engineering judgment compound instead of disappearing with the thread.
Subagents
Independent investigations run in parallel when the work genuinely splits: code tracing, test analysis, security review, or research. One plan still owns the outcome.
Proof surfaces
Static checks are only one layer. I separate local code proof, browser behavior, provider responses, data state, and production evidence so “done” means something.
Want an engineer who can direct agents, not just prompt them?
I bring the product judgment, systems thinking, and implementation discipline—and use a harness-independent AI system to make that leverage compound across the work.