Darren Card, fractional CPTOVancouver · 20 years in B2B SaaS
Building got cheap. Judgment did not. The operating model for a human-and-agent team.
The scarce work now is deciding what is worth building, and knowing whether what shipped actually landed. This site shows that work running, not a diagram of it.
Agents read, draft, and check. A person still makes every call.
On one team, “we got better” is a story. Across a dozen, read the same way, it becomes something you can see.
Twenty pages on what changes when building is cheap and judgment is not. Seven chapters, each listed with the claim it argues.
Read the contentsdid the onboarding revamp land?
Onboarding revamp
LandedActivation up 14 points six weeks after release.
MeasuredPricing page rebuild
WatchTrial starts flat. Sample is still too thin to call.
BenchmarkedIn-app messaging
StalledShipped nine weeks ago. Nobody has used it twice.
MeasuredOn one team, “we got better” is a story. Across twelve, read the same way, the teams shipping the most and landing the least are the easiest thing on the page to see.
The job did not change. One person still makes the call. What multiplied around them is the machine.
A dozen tools, four agents, and one owner. The interesting design question is not what the agents can do. It is where the line sits between their work and the call.
- 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.
Throughline runs this operating system on one real company. Vercel is the illustrative example.
One console, pointed at one company's public record, asking the same four questions a board asks. It is the only thing here that is built and running.
- Current
- 12% of shipped entries
- Healthy band
- 35–50%
- Norm
- 41% — this company's own 2023 baseline
- Modeled from
- entries_naming_outcome / total_entries
- Reads true if
- The classification pass was reviewed by a person this cycle and the changelog format has not changed. Both hold right now.
- Why did this move?
- What would move it most?
- Show me the inputs
Volume is not the problem. There are 3.1x more entries than two years ago. What changed is that entries naming a customer outcome fell from 41% to 12%, and the drop starts in the same quarter the release cadence doubled. The work is still shipping. The record of why it shipped stopped being written.
- entriesPublic changelog · 1,340 entries, 8 quarters
- outcome_classClassified against a written rubric, human-reviewed
- baselineSame changelog, first half of 2023
Every read is scoped to one subject and grounded in the model above. Point the same engine at a connected stack and the identical surface becomes a live read.Illustrative
Illustrative model of Vercel, built from public information.GET /api/v1/read
The read arrives ordered. Each move carries the metric it moves, the owner it routes to, and a projected band rather than a single confident number.
- 01
Overhead is already low, so the next improvement is moving review into evals
coordination overheadOperating modelRoutes to OperationsMove review into the eval gate and agent observability, not more process. The healthy band is near 8%.
Projected liftProjected+4.5 projected · band +3.5 to +5.5 · n=2 · last measured +4
Learned land rate75% · n=2
- 02
v0 to deploy attach is the drag on the revenue plan
v0 to deploy attachGo-to-marketRoutes to GTMTreat this as a product and pipeline problem together. The v0-to-deploy path is where generated apps become deploys, and where Sales and Marketing create and convert pipeline. Ship the path, set the attach target with GTM, instrument the ramp, and trace it to the v0 initiatives. Product builds the surface; GTM runs the motion on it; both stay true to the same number.
Projected liftProjected+5 projected · band −4.8 to +14.8 · n=1 · last measured +5
The band crosses zero, so this move could measure as a regression.
Learned land rate67% · n=1
- 03
A chosen bet is under-staffed
Enterprise trust, agent-safe allocationAllocationRoutes to OperationsThe owning team's call: move 5% of effort from Compute that scales to agents back to plan, restore Enterprise trust, agent-safe first this cycle, accept slower Compute that scales to agents progress as the cost, re-check next month. The drift just makes it hard to ignore.
Projected liftProjected+2 projected · band +0 to +4 · n=2 · last measured +3
Learned land rate50% · n=2
Illustrative model of Vercel, built from public information. The same deterministic engine the web read uses (deriveSignals over the operating model). No model invents a number. Numbers are derived, not asserted.
A working console, password gated. Opened on request.- 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
Watch it run on a real company
The console is password gated because the read is a working artifact, not a demo. Ask through the whitepaper form and I will open it with you. The reads above come from its public API, so the numbers are checkable without the password.
Vercel is the worked example because its record is public end to end. No affiliation, no endorsement. The reads above are the console’s own output, unedited.
The thesis gets tested in public, one short note at a time.
Ten notes, three minutes each, written from work in progress rather than from a finished position.
Agents broke the product org chart
Production got cheap. The chart we built around it stopped showing who's building the value.
Production got cheap. The chart we built around it stopped showing who's building the value.
- 09You can't ship AI you can't afford
- 08Evals are the new code review
- 07Velocity stopped meaning anything
- 06Hire for the PM job that exists now
- 05Smiling exhaustion is a system bug
- 04The one-person product org, from the inside
- 03The model was never the constraint
- 02Engineering is the only function you can see
- 01The CPTO is a consequence, not a trend