The Path Where the CEO Personally Masters AI Agents: Management with the Top Running 100 Agents Themselves [2026 Edition]

Hello, this is Hamamoto from TIMEWELL.

The leverage you get from "having staff use AI" versus "the CEO wielding it personally" is fundamentally different. In the first installment of this series, "Three Strategic Options for AI-Agent-First Management," we organized the three paths for layering AI agents onto management. This time I want to drill into one of them — the path where the CEO themselves runs the agents.

According to the Harvard Business School classic "How CEOs Manage Time" by Michael Porter and Nitin Nohria, who tracked 27 CEOs in 15-minute increments around the clock for three months, CEOs work 62.5 hours per week. Of that, 72% goes to meetings, and only 43% of their time is on agendas they themselves drive[^1]. The remaining 57% is consumed by "things that fall onto them." When I saw these numbers, I genuinely shuddered.

And yet, right now, we have at hand the tools to rebuild this from the ground up. Claude Code, Skills, MCP, Sub-agents. Run 100 agents in parallel and the very structure of the CEO's time changes. I want to write about both the implementation and the philosophy, including how I personally use them.

The gap between "CEOs who have staff use AI" and "CEOs who use it themselves"

In BCG's 2026 survey, 65% of CEOs ranked AI in their top three management priorities, and roughly three-quarters answered that "I am the final decision-maker on our AI strategy"[^2]. On the other hand, the National Bureau of Economic Research's survey of 6,000 executives that same year produced these numbers. Even among executives who use AI, average usage time is 1.5 hours per week. 69% use it less than one hour per week. About 90% answered that "AI has had no measurable impact on our productivity or employment"[^3].

What does this gap reveal when you read it calmly. Many CEOs have made the call to invest in AI, but do not personally touch it. Because they do not touch it, they have no gut sense of what the front line can achieve with AI. Because they have no gut sense, they cannot evaluate the quality of the output that comes back. They eventually summarize it as "we adopted AI but didn't see results." This is what is happening across a sizable share of Japanese companies — that is what I feel every time I talk to executives.

In a November 2025 interview, Google's Sundar Pichai said "the CEO's job may be one of the easier jobs for AI to replace"[^4]. Sam Altman has long said he would "welcome the day AI does my job better than I can." They can speak this way because they touch AI every day. Mark Zuckerberg has reportedly begun integrating an internal AI assistant into his own decision-making.

Let me state my stance here. The CEO of the future is the command center of AI agents. The CEO who only delegates to staff will probably be in a very tough spot five years from now. The reason is simple — you cannot evaluate the quality of how a weapon is used by an organization if you have never held the weapon yourself. A head chef who cannot grip a knife cannot run a kitchen. The gap between "executives who use AI and executives who do not" is no longer at the level of "executives who can type and executives who cannot." That is the starting point of this article.

What does the CEO delegate to AI agents

I do a weekly inventory of my own work, and the result always lands in roughly the same composition. Research for meeting prep is 25%. Drafting (proposals, approvals, IR, internal circulars) is 20%. Email and chat responses are 15%. Data checks and KPI reviews are 10%. Field and customer interactions are 20%. HR and management decisions are 10%. The first 70% is "work I want to delegate to AI" and the remaining 30% is "work I do myself."

Let me get more concrete about what is in that 70%. Track competitors' latest movements monthly. Pull stalled deals from your Salesforce and Notion. Detect KPI anomalies from financial data. Build out the market sizing of a new business from METI statistics and primary sources. Draft monthly shareholder reports. Write first drafts of customer replies. Compile a discussion-points memo for the next board meeting. Write Skills for these in Claude Code, run them in parallel with Sub-agents, and the results come back together by mid-morning.

First-pass email replies have outsized impact. In my case, I get around 80 emails a day. Six or seven out of ten are factual confirmations, scheduling, or light replies. I let an agent draft these via the Gmail MCP integration, and I press one of three options — "send," "edit," or "ignore" — in one second. That frees up about 15 hours a week. Redirecting those 15 hours into customer visits alone changes the revenue picture. Those are my own numbers.

Data analysis is also a place where agents shine. For example: "Of the 120 deals that came in through last month's inquiry form, give me the difference between those where we got a first meeting within 48 hours and those we didn't, broken down by industry and inquiry channel." Previously this required asking someone who could write SQL, waiting half a day, getting the result, asking follow-up questions, and waiting another half day. Now an agent runs the query directly against Supabase and returns the table in five minutes. The resolution of decision-making changed by an order of magnitude.

What I want to emphasize here is not to confuse "delegate" with "abdicate." The numbers an agent produces are ultimately read, internalized, and owned by the executive. To make this work, restrict to tools you can access yourself. Configurations where the agent can rewrite the entire company's data via MCP are dangerous. Clearly separate read-only and write-enabled, and always interpose human approval for writes. This is the first principle I share with clients of TIMEWELL's AI implementation service "WARP."

Redesigning the CEO's day: pouring time into meetings, the field, and customers

The Harvard finding that CEOs spend 72% of their time in meetings is what I call "the gravity of management." Meetings pull on the executive like gravity, and before you know it most of the day is gone. To resist this, the only option is to make almost everything that happens before and after meetings fully automated.

After much trial and error, I arrived at this kind of time design. Up at 5am. From 6 to 7, I read through the 30 reports that agents ran overnight. A research agent, a competitor monitoring agent, an internal KPI monitoring agent, an HR agent, an email drafting agent — each places its results in dedicated Slack channels. I read them in succession, give the necessary instructions, and reorder priorities. With this, 80% of the day's decisions are pre-staged.

7 to 8 is family time. From 8:30, meetings, field visits, and customer calls begin. From here, in principle, I do not touch AI. I have seen executives open laptops in sales meetings to consult AI as they speak — the other side cools off without fail. The CEO's field work is to deliver "your own words and judgment," and bringing AI in lowers the value of the product. I have decided AI is for use only before and after the field, never during.

From 17:00 to 19:00 is dialogue time with agents again. Pass the discussion points that came out of meetings to agents via Skills, with instructions like "give me three patterns of counterarguments to this point," "list ten weaknesses in this proposal," "narrow the next board meeting's agenda down to five points." Run these in parallel as Sub-agents and let them process overnight. By 6am the next day, 30 reports are waiting again. That is the cycle.

Behind this design is one goal: "Reduce work I personally do to under 10%." Of 100 tasks, 90 go to AI and 10 are done by my own brain. The 10 are final calls in meetings, sales closings, final interviews in hiring, steeling oneself for crisis response, dialogue with shareholders, and 1-on-1s with employees. These cannot and must not be handed to AI. Conversely, the company can run with all of the other 90 handed off to AI — or rather, designing the company so it does run that way is the executive's job.

To be honest, days where I pull this design off perfectly are still rare. Maybe half the month is a mix of old-style "the day I get swallowed by meetings." But the direction of the design is clear, and if I shift another 10% over to AI every month, I believe the landscape will look different three years from now.

The four foundations a CEO must master: Claude Code, Skills, MCP, Sub-agents

There are four foundations a CEO must master at minimum to run AI agents themselves. Claude Code, Skills, MCP, Sub-agents. In order.

Claude Code is the implementation of Claude that runs in the terminal. It does more than write text — it directly reads and writes files on your PC, runs commands, and connects to external APIs. If the executive is only touching "the ChatGPT chat box," the world is likely about to change here. I covered the details in "How to Use Claude Code Skills 4.5" and "The Complete Guide to the Superpowers Claude Code Plugin" — please read those alongside this piece.

Skills are like reusable instruction sheets. By Anthropic's own description, Skills are portable capabilities that span Claude Code, the Claude apps, and the API[^5]. For example: "When writing employee introductions, always include name, team, year of joining, and previous role at the top." "IR drafts must always include comparison numbers from the past six quarters." Write these company rules as Skills. You no longer need to communicate context every time — the moment a Skill is called, the form is already in place. The big deal is that "the way our company does things," which until now lived only in the executive's head, can be externalized as Skills.

MCP (Model Context Protocol) is a standard defined by Anthropic for connecting external tools to AI. Notion, Slack, Salesforce, Google Drive, GitHub, Figma, internal databases — they all connect through this. The value to the executive is simple: "AI can access all of your company's data directly." It can aggregate Salesforce deal data, Notion meeting notes, and Google Drive financial filings cross-functionally and write the report. Work that previously took a day by hand shrinks to ten minutes.

Sub-agents are a feature exclusive to Claude Code that runs specialist agents with independent contexts in parallel. For example, when you fire off a competitive analysis in one shot, you spin up a market research agent, a tech analysis agent, and a financial analysis agent simultaneously, each working independently. At the end, the main agent integrates the three results and hands them to the executive. You can run 100 or 200 in parallel — literally, the executive holds "an entire company's worth of research division" in their hands.

Combine these four and the number of agents one executive can run reaches, in theory, a thousand. As we touched on in "Google Cloud Next 2025: How AI Agents Are Changing the Way Enterprises Work," company-wide AI agent adoption is spreading rapidly, but examples of implementation at the individual executive level are still rare. That is precisely where the first-mover advantage is largest.

The "work only humans can do" the CEO must still own

I have spent this piece pushing "delegate to agents," but let me write the opposite at the end. There is work you must not hand to AI — or more accurately, work whose value drops the moment you do.

Final closing in sales. This is non-negotiable. What the customer's executive who signs off on a 5-million-yen purchase order is watching is not the contents of the proposal but "is this CEO trustworthy." A perfect proposal written by AI is beaten 10 to 20 percentage points in close rate by a clumsy verbal explanation from the CEO themselves. That is my own number. The existence of the executive is itself part of the product.

Final calls in HR. Who to hire, who to promote, who to put in charge of a new business. AI will say "based on past data, candidate A," but the air of the organization and the resolve of the person are not visible from where AI sits. The heavy decisions where misjudging shakes the entire organization should be made by a human executive willing to take responsibility. That is the structure that creates trust within the team.

Building relationships with shareholders and partners. This too cannot be handed to AI. Late-night dinners, a message asking after a family member's health, being the first to call when crisis hits — these build trust on a 10-year scale. You may send messages that AI drafted, but the final touch must be in the CEO's own words to mean anything.

Steeling oneself for crisis response. Compliance incidents, quality issues, natural disasters, talent walkouts. In moments like these, employees are watching "how did the executive decide." If you ask AI "how should we respond to this crisis," all that comes back is the lowest-common-denominator answer. The executive's resolve seeps not from data but from character.

Said inversely, the CEO's time should be concentrated on "work only humans can do." I often tell clients, "let's break down the executive's time budget." Out of 100 hours: 60 to field, customers, shareholders, employees; 20 to envisioning new businesses; 10 to risk and compliance calls; 10 to executing decisions. Research, drafting, and data tidying: zero. That is the right allocation. Once agents take over the several hundred hours of work behind the scenes, this 10:0 allocation becomes feasible for the first time.

In TIMEWELL's consulting service "WARP," we support this time design and AI agent implementation as a package for executives. Internally, we run "ZEROCK," an enterprise GraphRAG platform on AWS domestic servers in a closed environment, so you can hand confidential data to agents with peace of mind. If you want to connect the AI you personally run with internal data, this combination is the shortest route.

Three implementation steps a CEO can start today

I have written at length, so let me bring it back to action. Here are three steps you can begin tomorrow.

First. Install Claude Code on your MacBook and just touch it for an hour. It is fine even if you have never opened a terminal. Ask Claude Code "tell me how to use you" and it will teach you everything from the screen. In the first hour, if you can try the three things — "make it read a file," "have it draft an email," and "have it research the web" — that is enough.

Second. For one week, write out everything you do in 15-minute increments. Same method as the Harvard research. When you are done, look at the list and split it with two-color markers into "work I can hand to AI" and "work I do myself." Probably 70% of it will turn out to be work that AI can take. Just this discovery changes your sense of time as an executive.

Third. Write a dedicated Skill for the work you delegate. Start with one. In my case, the first Skill I wrote was "meeting summarizer." Hand it the audio of a meeting and it returns the discussion points, decisions, and action items in a fixed format. That alone freed up three hours a week. Then I added "competitor research," "KPI anomaly detection," "email drafting." After half a year I had about 20 Skills lined up, and they became my dedicated AI staff team.

Lastly, this is a personal feeling, but the time the CEO spends running AI agents themselves is far more enjoyable than you might imagine. Slowly escaping the gravity of "62.5 hours a week, 72% in meetings," the time you can spend thinking about what you really should be thinking about expands. Over the next few years, I think the class divide among CEOs will accelerate sharply. CEOs who do not use AI. CEOs who only delegate AI to staff. CEOs who run AI with their own hands. They will split into three layers, and productivity will differ by 10x at each level. Which layer do you want to be in? That is the only question.

---

References

[^1]: Porter, Michael E. and Nitin Nohria. "How CEOs Manage Time." Harvard Business Review, July 2018. https://hbr.org/2018/07/how-ceos-manage-time

[^2]: Boston Consulting Group. "As AI Investments Surge, CEOs Take the Lead." 2026. https://www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead

[^3]: Fortune. "Thousands of executives aren't seeing AI productivity boom." February 2026. https://fortune.com/2026/02/17/ai-productivity-paradox-ceo-study-robert-solow-information-technology-age/

[^4]: Fortune. "Google's Sundar Pichai says the job of CEO is one of the 'easier things' AI could soon replace." November 2025. https://fortune.com/2025/11/19/google-ceo-sundar-pichai-says-ai-can-do-his-job/

[^5]: Anthropic. "Skills explained: How Skills compares to prompts, Projects, MCP, and subagents." Claude Blog. https://claude.com/blog/skills-explained