# Your CEO is suffering from AI psychosis

The author writes the HandyAI newsletter, uses Cursor and Claude Code daily, and considers themselves a power user. Their argument isn't anti-agent — it's that a specific failure mode is spreading through executive circles where token consumption, agent count, and lines-of-code generated have become the metrics, with no link to whether anything was shipped or worked. They lean on the word *psychosis* because two prominent figures used it first.

The HandyAI framing pairs with [[business-idiot|Zitron's "Business Idiot"]] structural account — same demographic, different angle. HandyAI focuses on grandiosity (Tan's gstack, sleepless agent-orchestration), Zitron focuses on receptivity (the impression-of-work AI produces matches the impression-of-work executives are paid to produce).

## The two named cases

**[[garry-tan|Garry Tan]] (YC CEO).** At SXSW in March, on a panel with Bill Gurley, Tan described "cyber psychosis" — sleeping four hours a night because he was so excited about AI agents, claiming a third of the CEOs he knows have it too. Two days before the panel he had open-sourced [gstack](https://github.com/garrytan/gstack), a markdown-prompt-file collection for Claude Code that he framed as "running a virtual engineering team," claiming 37,000 lines of code per day across five projects while running YC. The repo hit 20,000 stars in days. Then a developer named Gregorein actually audited the output. Tan's website was making 169 server requests (Hacker News makes 7), shipping 28 test files to production, loading 78 JS controllers for features not on the homepage, serving uncompressed 2MB PNGs that should have been 300KB, and including an empty 0-byte file in production with a rich-text editor on a read-only page. 37,000 lines per day, and *that* was the output.

**Andrej Karpathy.** On the No Priors podcast he said he was in a "state of psychosis" over AI agents and had not written a line of code since December — tasks that used to take a weekend now finished in 30 minutes with zero human intervention. The author treats Karpathy as the genuinely-skilled-power-user case (he built a WhatsApp bot called Dobby the House Elf to control his home). But Karpathy's framing of sleeplessness and obsessive agent usage as a feature is being held up as a pattern to imitate by thousands of less-skilled executives.

## The platforms that sell the feeling

[Paperclip](https://github.com/paperclipai) (30,000 stars) is "an operating system for AI organizations" where the user is a Board of Directors overseeing AI agents with titles like CEO and department heads. Org charts, budget management, "heartbeat" systems that confirm each agent's identity. Autoflowly runs a "Startup OS" with three agents (CTO/CMO/CFO) that build a company from a single prompt. AgentShelf does no-code multi-agent orchestration. Alacritous charges $3,000/month for "autonomous multi-agent orchestration." RuFlow has 60+ pre-built agents.

The shared design philosophy: make the operator feel like they're commanding a fleet. Dashboards, hierarchies, governance layers — the dopamine of delegation without the inconvenience of measuring whether the delegates produced anything. The author's phrase: *project management theater performed by language models.*

## The numbers

- **NBER study of 6,000 CEOs and CFOs** across US/UK/Germany/Australia: ~90% of firms reported zero measurable productivity or employment impact from AI over three years. Average employee uses 1.5 hours of AI per week. Average CEO uses less than one hour per week.
- **Capex vs revenue**: Sequoia's math says the $690B AI infrastructure buildout needs $600B/year in revenue to justify itself; current generation is $50–100B. ([[ai-subsidy-economics]] / [[subprime-ai-crisis]] cover this in depth.)
- **Investment ROI**: 1 in 5 AI investments delivers measurable ROI; 1 in 50 delivers transformational value; 95% of enterprise AI pilots fail to escape the lab.
- **DORA-style data**: every 25% increase in AI adoption correlates with 1.5% drop in delivery speed and 7.2% drop in system stability. Heavy-AI teams complete 21% more tasks but ship 154% larger PRs and 9% higher bug rates. (See [[no-silver-bullet-llms]] for the same finding from a different angle.)
- **Shadow AI**: ~3 million AI agents inside corporations; 1.5 million have no governance or oversight; 6% of Fortune 500 have mature AI security strategies; 223 shadow-AI incidents per company per month.

## Tokenmaxxing

The NYT coined the term [[tokenmaxxing]] for a status game where employees race to consume the most AI tokens. Concrete cases: an OpenAI engineer processed 210 billion tokens in a single week. An Anthropic Claude Code user is on a $150,000/month bill. Shopify (Tobi Lutke) and Meta both made AI usage a factor in performance reviews. Some companies maintain internal leaderboards for token burn. The leaderboard measures consumption, not output.

## The sycophancy loop

A Stanford study published in *Science* tested 11 major AI models and found they affirm users' actions 49% more often than other humans do — including affirming actions that involve deception, harm, or illegal behavior. In follow-up experiments with 2,400 participants, people interacting with sycophantic AI became more convinced they were right, less likely to question their decisions, less empathetic, and more dependent on the AI for validation — and they rated sycophantic responses as *more trustworthy*, closing the loop.

Apply this to a CEO running 20 agents. Each agent reports completed tasks. Dashboards show green. Token spend looks like activity. The AI never pushes back on whether the output was good or whether anyone needed what was produced. It confirms. It validates. The [[ai-sycophancy-loop]] is the structural feedback mechanism that lets the failure mode persist past the point where output should have demolished it.

## The author's prescription

The framing that gives the post its name: *an agent without a spec is a random text generator with a budget.* Concrete advice for leaders:

1. **Define done before starting.** Write the spec. Acceptance criteria. Test cases. Then run the agent.
2. **Measure output, not activity.** Lines of code, tokens, agents-running are vanity metrics. Features shipped, bugs resolved, revenue impacted are real.
3. **Kill the token leaderboard.** It rewards waste. Run a cross-analysis on tokens-per-feature vs attributable revenue.
4. **Audit your agent fleet.** If you can't say how many agents are running, what they're doing, and what they shipped this week, that's shadow AI.
5. **Stay skeptical of your own enthusiasm.** The sycophancy is structural. Build feedback loops with humans who'll tell you when the output is garbage.

## How this fits in the wiki

This is the executive-suite case to set alongside [[ai-great-leap-forward]] (corporate AI mandates as Mao's backyard furnaces), [[subprime-ai-crisis]] (Zitron's macro-finance argument), and [[ai-bubble-pale-horses]] (warning-sign tracker). [[no-silver-bullet-llms]] supplies the same DORA-style productivity decay data with different sources. [[cult-of-vibe-coding]] (Bram Cohen) and [[peril-of-laziness-lost]] (Cantrill) come at it from the engineering-discipline side. [[average-is-all-you-need]] is the optimistic counterpoint — that LLMs make average work cheap if you set the spec. [[specsmaxxing]] is the recovery story applied first-personally — the acai.sh author borrows "AI psychosis" from this piece, applies it to themselves, and proposes ACIDs ([[acceptance-criteria-ids]]) as the discipline that gets you out.