Don't give Fable all the powerBlog/2026-07-04/Silver

Don't give Fable all the power

Why the smartest model made the worst orchestrator. Notes from a week of running an autonomous AI firm.

I spent the last week running a fully autonomous multi-agent system: a little firm of 20+ LLM agents doing empirical research around the clock on my desktop. The domain does not matter for this story. What matters is that wrong answers look exactly like right ones until you test them, so the harness is strict: holdout evaluations, placebo tests, multiple-testing accounting. The first thing you learn is that your agents will happily overfit their way to fictional breakthroughs.

The obvious architecture, the one every tutorial suggests, is to put your smartest model at the top. Fable, Opus, whatever your frontier tier is: make it the orchestrator and let it delegate to cheaper workers.

I ran that. It failed in three distinct ways.

Where the big-brain orchestrator breaks

1. Context saturation. The orchestrator accumulates every worker's report. Once its context gets long, it starts forgetting its own earlier decisions and re-issuing work it already dispatched. You get loops, not progress.

2. Cost gravity. When every orchestration decision costs frontier-model money, you start designing for fewer, bigger ticks without quite noticing you are doing it. The swarm idles between expensive deliberations. Utilization was terrible.

3. The conclusion instinct. This one surprised me. A frontier model, handed a mountain of evidence, does what it is trained to do: synthesize and conclude. Twice I found the system parked overnight because the orchestrator had written a genuinely well-argued memo ("all research avenues exhausted, recommend stopping") and stopped. The memo was even right in a narrow sense, and that is exactly the problem. Research does not stop, it reframes. Intelligence at the top gives your system an off switch, and the eloquence to justify pressing it.

The stupid-CEO pattern

What fixed it:

The orchestrator is a cheap model, and it is content-blind by design. It never reads the research. It reads one fixed dashboard (a shell script emitting JSON: open task count, staleness, queue depth), refills empty worker slots, and reschedules its own wake-up. It cannot conclude anything because it cannot understand anything. That is not a compromise. It is the load-bearing feature.

Mid-tier PMs are stateless. Fresh context every tick: read state from files, dispatch workers, write state back, terminate. No agent lives long enough to accumulate a worldview.

The frontier model works in the basement, not the corner office. Root-cause analysis, adversarial audits of results, strategy reviews: short, judgment-dense sessions on curated context. Then it is gone. It never holds the steering wheel.

All memory is files. One file is the single source of truth for the current best result. Append-only lesson logs record what has been tried. A registry of attack codes stops agents from re-running ideas that already died. A janitor agent trims everything against hard size budgets, because these files otherwise balloon to megabytes in days.

Verification is mechanical, not vibes. Every claimed result has to reproduce through a gated replication run with placebo controls before it can touch the record. LLM enthusiasm dies at that gate constantly. That is the gate working.

One iron rule: never an empty tick. The scheduler cannot rationalize idling, because it cannot rationalize anything.

Results, honestly stated

The system now runs unattended for stretches of 8 hours and more. It has opened and killed 600+ research directions, with documented statistical evidence for each kill. Its best validated result came after the flip to the stupid CEO. And the most interesting output increasingly comes from the kills: a map of what does not work, with receipts.

Whether the research itself ever hits its targets is a separate question, and I am much less sure about that one. But the systems lesson feels solid, and it transfers.

Capability and control want to be different axes. Put judgment in short-lived, context-curated bursts. Put persistence in a loop too dumb to stop.

The org chart had it right all along. The genius belongs in the lab, and the thing ringing the bell every twenty minutes should barely know what the company does.<br>Sell it before you build it.<br>Real ads. A real page. Real buyers in 3 days. Your AI team turns that proof into a product built to sell.

Get started