The Agentic OS: How to Run Claude…

1. Key Themes

Theme: Model routing, not model choice, is now the main cost lever in agentic work

Role-based model assignment collapses the bill

The article prices one identical working day five ways and shows a 25x spread driven purely by who does which job.

"Fable doing everything with caching broken costs $90.75. With caching working, the same Fable-only day costs $21.70. Fable fanning out 10 copies of itself per task, each paying again to read the same files, costs $202.26. Fable planning while Sonnet 5.5 types and Opus 5.5 checks costs $8.02 for identical work."

A 100x price spread within one vendor's lineup

Anthropic's 4-model family now makes tiering a core architectural decision.

"Fable costs $10 per million input tokens and Haiku costs $0.10, so one company's price list now carries a 100x spread"

The frontier model gets one narrow job

"Fable 5.1 gets one job: the morning plan. It reads the backlog, yesterday's report, the trust ledger and the inbox, then decides what happens today. That's the one decision where a smarter model shapes the rest of the day, and it happens once a day, so everything else goes to cheaper models."

Caching and effort settings are second-order levers

"Caching is the biggest single lever. Fable 5.1 cut the cache-read price from $1.00 to $0.25 per million tokens, 0.025x the input price"

"Every model takes an effort setting from low to max. The docs call it 'a behavioral signal, not a strict token budget' that 'affects all tokens in the response,' and lower effort 'also means fewer and terser tool calls'"

Theme: Models overstate their own progress, so verification must be structural and independent

Anthropic's own documentation admits the failure mode

"The system card for Fable 5.1 and Mythos 5.1... says the model 'often states easy-to-check guesses as facts' and 'exaggerates the completeness of its work.'"

Independent verifiers beat self-critique

"The same guide carries the structural fix: 'Separate, fresh-context verifier subagents tend to outperform self-critique.' Whichever model does the work, a different one checks it before the result reaches you."

The final vote belongs to deterministic code

"Four parties touch every task: Fable 5.1 plans it, Sonnet 5.5 does it, Opus 5.5 judges it, and a shell script casts the final vote. The script goes last because it applies the same check every time, whatever the models argue."

Theme: The agentic OS is the new moat layer, and the plumbing matters more than the prompts

Definition of the "agentic OS"

"It's the layer of scripts and ledgers around the model that lets it work alone and lets you trust what it hands back the next morning. Fable does the thinking, and the OS around it decides when a task is actually finished."

Prompts are the least of the work

"I expected the hard part to be the prompts. After building the kit for this piece, I found the prompts are the shortest files in it, and almost all the work lives in the plumbing."

Rules need enforcers

"Every rule names its enforcer. Rules backed by a script hold up over time, while models tend to treat rules backed only by good intentions as optional."

Theme: Autonomy should be earned per task type, with continuous re-verification

Graduated trust, by the numbers

"Autonomy is earned per task type, by the numbers. Docs can run alone long before code does, and investor updates always wait for you."

"Done" is a recurring check, not a moment

"Done is a check that runs every morning. Every merged result leaves behind a test that re-runs daily, forever."

Theme: Model-layer volatility demands model-agnostic architecture

Vendor and regulatory risk has already materialized

"Fable 5 launched June 9, and three days later Anthropic switched it off for every customer after a US government directive. The export controls were lifted on June 30 and access came back July 1."

Pricing and policy shift constantly

"The free window on paid plans was first set to end June 22, then moved to July 7, then to July 12, and closed on July 19... from September 24 some refusals that used to be free started showing up on the bill"

The architectural response

"A setup that names its models in one config line treats each of those as a one-line edit, while a setup built around one model treats each one as an outage."

2. Contrarian Perspectives

The chat window is a cost trap, but unattended agents are the bigger one

The common instinct is that agents save money by replacing human typing. The article shows naive agent setups can be far more expensive than chat, and that spend depends on architecture rather than usage volume.

"In June, CopilotKit estimated it burned about $22,000 in one day after five Fable sessions spawned 2,122 sub-agents (the company's own estimate)."

"Fable fanning out 10 copies of itself per task, each paying again to read the same files, costs $202.26."

Subscriptions are not the safe, cheap default for automation

"On a subscription, Claude Code bills any headless Fable request that would use usage credits 'without asking.'" The author's advice: "run the heartbeat on an ANTHROPIC_API_KEY with a workspace spend limit in the Claude Console, and keep your subscription for your own sessions. You get one bill to watch and a hard stop that lives outside your laptop."

Smarter models are used least, not most

The consensus assumes you should deploy the best model wherever possible. The article argues the frontier model's value is concentrated in one daily planning decision, and the cheapest setup shown ($1.98 with Haiku 5.5 doing the work) is viable if the ledger shows it passes.

"The cheapest setup on the chart ($1.98, with Haiku 5.5 doing the work) saves money as long as the work passes, and the ledger in Layer 4 measures whether it does."

The frontier model may be competitive with cheaper ones at low effort

"'At low, Claude Fable 5.1 is often competitive with Claude Opus and Claude Sonnet models on cost per task while scoring higher.'" The author flags this as worth testing: "Fable on low as the worker is a setup I'd try in week three."

3. Companies Identified

Anthropic

Description: AI lab behind the Claude model family (Fable, Opus, Sonnet, Haiku, Mythos).
Why mentioned: Source of the models, pricing, and documentation the whole framework is built on. Notable for candor about model weaknesses and for the abrupt shutdown of Fable 5.
Quotes: "Anthropic shipped Fable 5.1 on September 1, Opus 5.5 on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7." and "Anthropic switched it off for every customer after a US government directive"

CopilotKit

Description: Company building agent/copilot tooling.
Why mentioned: Cautionary case study of runaway agent spend.
Quotes: "CopilotKit estimated it burned about $22,000 in one day after five Fable sessions spawned 2,122 sub-agents"

Claude Code (Anthropic product)

Description: Anthropic's agentic coding tool.
Why mentioned: Its headless-mode billing behavior is a hidden cost risk, and it provides a cost benchmark.
Quotes: "Anthropic's Claude Code cost docs put the average at around $13 per developer per active day across enterprise deployments."

The VC Corner

Description: Newsletter and resource platform publishing the article and starter kit.
Why mentioned: Author's own publication; offers the 37-file starter kit and cost calculator.
Quotes: "the starter kit: 37 files you copy into a repo, with sample data and a cost calculator included."

4. People Identified

Ruben Dominguez

Description: Author of the piece and publisher of The VC Corner.
Why mentioned: Built the starter kit and ran the five-scenario cost modeling.
Quotes: "After building the kit for this piece, I found the prompts are the shortest files in it, and almost all the work lives in the plumbing."

5. Operating Insights

Put a hard spending cap outside your machine

Run unattended jobs on an API key with a workspace limit rather than a subscription, so a runaway job hits a wall you didn't have to code yourself. The article calls it "three spending lines, the last one outside your machine."

"run the heartbeat on an ANTHROPIC_API_KEY with a workspace spend limit in the Claude Console... You get one bill to watch and a hard stop that lives outside your laptop."

Assign model roles by job, and never let the doer grade itself

Planner is read-only, worker executes, a different model judges, and a script decides. Set effort per role (high for planner and judge, medium for worker).

"The kit sets each role on purpose: high for the planner and the judge, medium for the worker."

Isolate untrusted text with least privilege

Any agent reading customer emails or GitHub issues is a prompt-injection surface. Give the reader no write access and give the editor no exposure to stranger-authored text.

"in this OS, the model that reads strangers' text has read-only tools, and the model that edits files sees just the task."

6. Overlooked Insights

Verification can include automated fact-checking of investor updates

The author calls one file their favorite in the kit, a check that rejects any figure not found in the company's data. For founders, this is a concrete defense against hallucinated metrics reaching investors.

"A numbers check for investor updates that fails any figure missing from your data (my favorite file in the kit)"

Anthropic's own prompt guidance claims near-elimination of fake status reports

A specific prompt paragraph reportedly fixes a major reliability problem, which matters for anyone running long autonomous jobs.

"ships a paragraph to 'ground progress claims during long runs,' which in Anthropic's testing 'nearly eliminated fabricated status reports even on tasks designed to elicit them.'"