My COO Is an AI
I structured my AI assistant like an organization: a chief of staff, a council of advisors, and domain experts. It sounds absurd. It works.
I experiment with technology on weekends. Multiple explorations running in parallel, each at a different stage. Keeping track of what I decided, why, and what's stalled across all of them was becoming impossible.
So I did what felt natural after years of managing a large account with multiple stakeholders: I gave my AI assistant an organizational structure.
One point of contact
Most people use AI as a question-and-answer tool. You ask, it responds, you move on. I wanted something different: a single AI that I talk to, that knows all my explorations, and that coordinates with specialized agents and advisors when the question is bigger than what it can handle alone.
I talk to one agent. It decides who else needs to be involved. Sometimes the answer is nobody. Sometimes it convenes advisors with different perspectives. Sometimes it pulls in a specialist: a financial one to pressure-test a pricing idea, a legal one to read a contract before I sign it. But I never have to think about routing. One conversation, one point of contact.
The structure
Me: Direction, priorities, judgment calls. I decide what to explore and why.
The AI (chief of staff role): Cross-exploration awareness, proactive alerts, institutional memory, operational coordination. It tracks whether things are progressing, what's falling through the cracks, and what the data says about decisions I've already made.
The title depends on which way it faces. Toward me, it's a chief of staff. It advises, remembers, and needs my yes on anything that matters. Toward the system, it's a COO. It runs the operation: who gets involved, what happens next, whether it actually got done.
The advisors: two specialized seats with different thinking styles. The Theorist grounds discussions in research and frameworks. The Validator builds evaluation frameworks with measurable criteria. They get convened only when a question is bigger than the chief of staff should decide alone.
The advisors respond independently, so neither anchors on the other. Then the chief of staff synthesizes, notes where they disagree, and gives its own recommendation. One agent coordinating specialists, not a committee.
How a decision actually flows
I'd been noticing that my AI kept forgetting decisions between sessions, repeating mistakes I'd already corrected, losing the reasoning behind choices. I asked the chief of staff: "Should we build a proper memory system, or just accept the amnesia and re-explain things each time?"
That is not a question it answers alone. It touches every exploration. The fix costs a weekend. And being wrong is expensive in a way that is hard to see. So it escalated.
Step 1: The chief of staff audits every exploration and comes back with numbers. Nine of ten had no persistent knowledge files. The session log had grown to 6,800 lines that nobody read. I had 22 files of corrections I'd given it. Eleven had never been loaded into a session.
Step 2: Convenes both advisors on the same brief, audit attached, neither one seeing the other's answer.
Step 3: The Theorist maps the problem onto the Atkinson-Shiffrin memory model from cognitive psychology. Session logs are raw sensory input that decays fast. What's missing is the encoding step: filtering raw input into something retrievable. Recommends a centralized semantic index with federated context stores.
Step 4: The Validator builds a measurement framework. Six metrics including recall accuracy, citation rate, and false memory rate. Defines concrete "done" criteria for each phase: "100% feedback files indexed, decisions loaded at startup, baseline dashboard live." Specifies failure modes and how to detect them.
Step 5: The chief of staff synthesizes. Both advisors agree the problem is retrieval, not storage. Recommends a phased build starting with the highest-value files, those 22 corrections. Proposes a 60-question evaluation suite to measure improvement.
Step 6: I decide. Agree with the phased approach. Build starts that weekend. The result was dramatic.
Total time for me: about 15 minutes of reading and deciding. The system did the research, the framework building, and the synthesis. I did the judgment.
The compound effect
That session didn't just produce a decision. It produced dated minutes: raw advisor opinions preserved, synthesis kept separate, decisions and action items explicit. Those minutes became the spec for the build, then searchable context for every decision after it. This is compound engineering in practice: each session makes the next one smarter. Decision reasoning that used to survive about 40% of the time now persists by default.
It compounds mistakes too. A few weeks ago the system decided one of my machines had no working GitHub credentials. It had. The check had run in a shell that couldn't see the tool it was looking for. Nobody re-tested it. That wrong fact then loaded at every session start for two weeks. It quietly ruled out a machine I could have been using the whole time. Memory that spans sessions propagates errors as faithfully as it propagates good decisions. The fix was a rule: never write down an absence you haven't actually verified. That rule now loads at session start too.
Such an elaborate organization for weekend tech musings sounds absurd. But it works.
What this is not
This is not a chatbot with a fancy title. A chatbot answers questions when asked. My system reads context proactively, notices when things stall, flags forgotten follow-ups, and connects patterns across explorations. A chatbot forgets when you close the tab. This system's memory spans sessions through versioned knowledge files.
This is also not "AI replacing thinking." The system produces zero value without me making decisions and exercising judgment. What it replaces is the cognitive overhead of keeping track of everything.
What's changed since
This post described the org chart in March. By June, the same structure had grown limbs.
The two advisors became a seven-seat council of different frontier models. Different models disagree in more useful ways than different prompts ever did. Two prompts against one model fail the same way at the same time. Two different models fail differently. The disagreement is where the information is. Same brief, seven independent answers, in one of three modes depending on the question: a single parallel round, a two-round debate where seats challenge each other, or role-play with assigned hats.
The chief of staff now runs an autonomous loop that picks tasks off a shared kanban board while I sleep, inside a strict safety envelope. There is a voice interface, so some of those board cards get created from my couch.
The org chart held. The headcount didn't. Here's what six months of that actually looks like.
I didn't set out to build an org chart for my AI. I set out to stop forgetting things. The structure emerged because it was useful.