Tools
Each tool answers one question about your own situation. The browser tools are free and need no login, and what you type in them stays in your browser.
The live bench
Pick a tool and answer its first question; the full tool goes further.
Which layer does your AI advantage run through?
The thing a new AI capability has to pass through to reach your customer.
Pick a layer
Picture one task you might split across agents.
Can each part be finished without knowing what the other parts decided?
Do the parts only read, or write into one shared result?
Is it too big for one context window?
Is it worth roughly 4× what one agent would cost?
Answer the four
How reliable is each step your agent takes?
Per-step success compounds over a long run. Slide the three dials.
Which number do you steer by?
Pick the closest. The reading is how that kind of number misleads you, and the check that catches it.
Markets & investing
SaaS & growth
AI & infrastructure
Operations & engineering
Pick a metric
Pick one automated loop you rely on.
Could the signal be faked?
Is the first failure terminal?
Answer both
What kind of belief about your numbers is it?
Pick the kind
How fast does your number grow each period?
A compounding curve starts out looking straight. This is how far it has to grow before the test can tell.
How it works The numbers and verdicts come from each tool’s own code, copied from its page when this site is built; the bench adds the layout and the sentence around them. Nothing you pick here is stored or sent.
Whether the advantage you have is still paying, and for how long.
You getRoughly when your AI advantage stops paying, and how soon to start building the next layer.
Two steps · about a minute Stores nothing
From the essay How Long Until Your AI Edge Stops Paying?You getWhether your curve is compounding or growing in a straight line, and how late you can tell.
Build a curve, bring your own, or run it blind Stores nothing
To print or download
The PDFs are free on Gumroad, which asks for an email. The downloads run on your own machine.
The Displacement-Rate Audit
Get the PDF: The Displacement-Rate Audit, free on Gumroad, opens in a new tab
The Game Check
Get the PDF: The Game Check, free on Gumroad, opens in a new tab From the essay behind The Game Check
Whether the system you built does the job you think it does.
You getA call on each task, ranked: one agent, a hybrid, or a team, and the answers that decided it.
Your tasks, then four questions on each Saves in your browser only
From the essay Your Multi-Agent System Is an Org ChartYou getHow often a long agent job actually finishes, and what a finished task really costs.
Two models, your job length, what a run costs Stores nothing
From the essay Your Benchmark Measures a Sprint. Your Agent Runs a Marathon.To print or download
The PDFs are free on Gumroad, which asks for an email. The downloads run on your own machine.
The Seven-Layer Agent Audit
Get the PDF: The Seven-Layer Agent Audit, free on Gumroad, opens in a new tab From the essay behind The Seven-Layer Agent Audit
The Skill Bill of Materials
Get the PDF: The Skill Bill of Materials, free on Gumroad, opens in a new tab From the essay behind The Skill Bill of Materials
The Consolidation Test
Get the PDF: The Consolidation Test, free on Gumroad, opens in a new tab From the essay behind The Consolidation Test
The Skill Routing Eval
Which of two competing skills Claude actually fires, graded over twelve labelled prompts.
Download The Skill Routing Eval, 11 KB How it works: The Skill Routing Eval
The State Channel Audit
Where a parallel step can silently overwrite a decision in your own LangGraph schema.
Download The State Channel Audit, 25 KB How it works: The State Channel Audit Printable worksheet, a 4-page PDF, 349 KB
Whether the number, test or check you trust is telling the truth.
You getThe specific way your number can mislead you, and the check that catches it.
Choose one metric, then answer its audit Stores nothing
From the essay The Stable LiarYou getEach loop placed on a map, and the move for any in the risky corner.
Two questions per loop Saves in your browser only
From the essay Your AI Looks Best Where You Can Check It LeastYou getFor each belief, a verdict on the trap it may hide (a false trade-off, a safety net that fails on the worst day, parts that drag each other along) and the fix.
Per belief: its kind, then one or two questions Saves in your browser only
From the essay Your AI Stack Has Three Bugs Other Fields Already FixedYou getWhether a robot number counts robots ordered, built, shipped, installed or working, and what to ask the company next.
Five questions · two minutes Stores nothing
From the essay Your Robot Coworker Is Still a PilotTo print or download
The PDFs are free on Gumroad, which asks for an email. The downloads run on your own machine.
The Looks-Right Trap
Get the PDF: The Looks-Right Trap, free on Gumroad, opens in a new tab From the essay behind The Looks-Right Trap
Tests Worth Passing
Which of your tests notice when the code under them is broken on purpose.
Download Tests Worth Passing, 13 KB Read the essay: Tests Worth Passing
