Eight beatsSix minutes end to end
Building got cheap. Deciding what to build, and knowing whether it landed, did not.
Eight claims in order. By the end you should be able to tell whether your own team is counting effort or counting outcomes, and what it would take to change that.
Claim 06 is the one that changes a review the following Monday. Everything before it explains why it works, and everything after it shows the cost of skipping it.
Jump to claim 06The expensive part of software stopped being the writing.
An agent now does in an afternoon what took a competent team two weeks. I have done it myself and watched other teams do it. The part that did not get faster is the part where somebody looks at the result and decides whether it is right.
Nine days of drafting, reading, and checking. One day of deciding.
The drafting collapsed. The deciding is exactly as long as it was.
Two things are scarce now: deciding, and knowing whether it worked.
When capacity was the constraint, prioritising was mostly triage. Now that almost anything can be built this week, the choice itself is the expensive act, and so is finding out whether the choice was right.
- Writing the code
- Reading the tickets
- Drafting the options
- Checking the rules
Struck through because these stopped being where the cost sits, not because they stopped mattering.
Deciding what is worth building
When you can build almost anything this week, choosing becomes the expensive act.
Knowing whether it worked
Shipping is visible by default. Landing is only visible if somebody sets out to see it.
People decide and judge. Agents build. Five steps, repeating.
This is the whole operating model. It is not a methodology and it does not need a rollout. It needs one owner per step and the discipline to finish step five before starting step one again.
- 01Person
Decide what is worth building
A named person picks the bet and writes down what would make it a win.
- 02Agents
Read everything that touches it
Tickets, calls, deploys, and past attempts, pulled together in minutes.
- 03Agents
Draft it, build it, check it
Options written, code produced, rules and edge cases tested against.
- 04Person
Approve it or send it back
One owner signs off on business fit, design quality, and technical fit.
- 05Person
Judge whether it landed
Read the same way every time, so the answer means the same thing twice.
- Then step 01 again
What you learn in step 05 is the only honest input to the next decision. Skip it and you are guessing, faster.
Agents never decide. Every call has one owner with a name.
Agents are extremely good at producing defensible-looking work, which is exactly why the line has to be drawn explicitly. A recommendation is not a decision until a person puts their name on it.
- Business fit
- Design quality
- Technical fit
- Final approval
Judgment. Never delegated, never automated.
- Read the context
- Draft the options
- Check the rules
- Propose the next move
Volume. Fast, tireless, and never the owner.
Everything left of the rule has a name attached to it.
Ordered by expected effect on how the team works, not by age or by who asked. Each one arrives already dry-run against real history, so the decision is a yes or a no rather than a research project.
- 01
Require an outcome line in every changelog entry
Propensity0.86What it changesAdds one required field to the release template and blocks publish when it is empty.
Dry runReplayed against the last 90 entries. 79 would have passed untouched; 11 needed a single sentence.
Awaiting DarrenLog decisionPush to Linear - 02
Kill the Thursday portfolio review, replace it with a written read
Propensity0.71What it changesRemoves a recurring twelve-person meeting and adds one written read due Wednesday.
Dry runSix of the last eight sessions produced no recorded decision. Those six become the written read.
Awaiting DarrenLog decisionPush to Linear - 03
Move the notification centre to Stalled and free the two engineers
Propensity0.64What it changesReclassifies one initiative and returns its headcount to the next cycle.
Dry runLive for two quarters with no measurable change in any account. Nothing downstream depends on it.
Awaiting DarrenLog decisionPush to Linear
The engine wrote, scoped, and dry-ran all three before anyone opened this screen. It cannot approve one, and that is the whole design.Illustrative
The work lives in a dozen tools. The answers live on one screen.
Nobody is going to consolidate their stack, and they should not have to. What has to be consolidated is the reading: one place where the same three questions get asked of every launch.
- Tickets
- Deploys
- Analytics
- Support
- Calls
- Billing
- Docs
- Experiments
- Reviews
- Surveys
- Incidents
- Roadmap
- Did it land
- Who owns the call
- What to do next
Same three questions every week, asked the same way.
- ChangelogPublic release notesFresh · 12m ago
What actually shipped, when, and whether the entry names a customer outcome or only a mechanism
- entries
- ship_date
- outcome_class
- Docs & pricingPublic product surfaceFresh · 1h ago
How many surfaces exist to learn, and which ones changed shape between releases
- surface_count
- pricing_change
- Job postingsPublic org signalFresh · 6h ago
Where headcount is actually being pointed, which is the most honest statement of priority a company publishes
- role_mix
- org_signal
- Investor commentaryPublic earnings recordStale · 3d ago
What leadership said the priorities were, in their own words, on the record
- stated_priority
- What they decided to buildentries grouped by surface + stated_priority
- Which bets show they landedoutcome_class over trailing 2 quarters
- Where effort and outcome splitrole_mix vs outcome_class by surface
- What the next call probably isranked by gap x reversibility
Nothing is inferred from a source that is not listed here. When one goes stale it is marked, and the reads that lean on it inherit the mark rather than being quietly served as current.Illustrative
Stop counting what shipped. Count what landed.
Effort is easy to count, so it becomes the report. Outcome is harder to count, so it becomes an anecdote. Swap which one you count and the conversation changes inside a week.
Landed, watch, or stalled. Three words, applied to everything, including the work you personally championed.
- Onboarding revamp14 itemsNew teams reach their first invite in a day, not a weekLanded
- Billing self-serve9 itemsPlan changes stopped arriving as support ticketsLanded
- Search rebuild22 itemsUsed more, abandoned at the same rate. Two more weeks of readingWatch
- Mobile parity31 itemsShipped in full. Nobody has changed what they do yetWatch
- Notification centre18 itemsLive for a quarter. No measurable change in any accountStalled
- Workspace roles12 itemsBuilt for an objection that no buyer actually raisedStalled
Illustrative model of Vercel, built from public information.GET /api/v1/landed
Both of these shipped. The verdict is about whether behaviour moved, which is a different question and the only one that pays.
- Watch
Lift new-user activation
activation rateBaseline47%Target60%Landing read fromauthoredThe onboarding revamp is working: activation has climbed to 54% from 47% and is trending toward the 60% bar. The top growth lever, and it is moving in the right direction.
- Watch
Keep spend predictable as agent usage scales
Predictable spendBaseline62%Target90%Landing read fromauthoredCost per agent workload is the lever; surprise usage bills are the failure mode.
Illustrative model of Vercel, built from public information. The same deterministic engine the web read uses (deriveLanding over the operating model). No model invents a number. Numbers are derived, not asserted.
A working console, password gated. Opened on request.A working thesis on human-and-agent teams, and the system that runs it.
The argument and the instrument are the same project. One of them is written down and one of them is running, and each keeps the other honest.
A working argument, not a finished one
Twenty pages on what changes when building is cheap and judgment is not. Written to be disagreed with.
The same argument, running on a real company
Throughline asks the four board questions against one company’s public record, the same way every week.
The alternative is guessing faster.
Agent tooling makes a team quicker at producing work. On its own it makes them quicker at producing the wrong work, and the spend line keeps climbing either way.
The two lines used to move together. They stopped.
Now watch it run on a real company.
Throughline points this operating model at one company's public record and asks the four questions a board asks. The console itself is password gated, and the homepage beat shows exactly what it produces.