Skip to content

Built in public · iterates most weeks · last shipped 2026-08-22

dacard.ai
The point of view

Twenty years, revised in publicThe thesis is the long version

Shipping totals stopped telling you who is getting better.

The scarce work moved to deciding what is worth building and knowing whether it landed, including the parts that make a quarter look worse than it did before.

If you read one thing

Principle 2.2. Agreeing what landed means before the work starts is the step almost nobody takes.

Four principles01
Held in public

Four things I will say out loud, and change my mind about if the evidence turns.

Each one is testable, and 2.4 is the limit of the other three.

2.1

Shipping totals stopped telling you anything.

When production capacity rises sharply, output stops separating a team that is getting better from one that is getting busier.

2.2

Landed is a definition, agreed before the work starts.

A behaviour, in named accounts, by a named date. Set beforehand, it survives the quarter going badly.

2.3

Agents produce. One person owns every call.

Reading, drafting and checking move to agents on a schedule. What to build, what to decline and what to stop stays with somebody who has a name.

2.4

Measuring landing without doing discovery just makes teams better at defending what they shipped.

Score a team on whether work landed and they will optimise for defensible launches: safe, small, easy to attribute. The counterweight is discovery. What to build next comes from customers, and the definition of landed is set with the people measured by it.

The operating model02
Person and agents

How a product team runs when agents do the producing.

One named owner on every call, and no piece of work closes until somebody checks the number it was meant to move. A dozen tools, several agents, one owner. The design question is where the line sits.

2015One person. One screen.
One person, deciding
2026One person. A wall of interfaces, agents alongside.
One person, deciding
The number of surfaces a product decision touches went up by an order of magnitude. The number of people who own the decision stayed at one.Illustrative
Person
  • Business fit
  • Design quality
  • Technical fit
  • Final approval

Judgment. Never delegated, never automated.

The call
Agents
  • Read the context
  • Draft the options
  • Check the rules
  • Propose the next move

Volume. Fast, tireless, and never the owner.

Everything above the rule has a name attached to it.

How it runs03
Cadence, rights, and the one artefact

A model you cannot put in a calendar is a philosophy.

Three rhythms, six calls with a name on each, and the brief that starts everything. The rest of the model is in the thesis; these are the parts a team needs before Monday.

The calendar
  1. WeeklyThe read

    Nine channels, scanned in the order they always appear, and a staleness sweep, because context rots quietly and nothing raises an exception when it does. One person runs it and anyone can read it without asking them.

  2. Every cycleThe verdict

    Each shipped thing gets one of three words against the bar agreed before it started. Landed, watch, or stalled. A stalled costs something to say and is the only one that changes what happens next.

  3. QuarterlyThe readout

    Four answers, the portfolio underneath them, and one decision carried in with a number attached. What it says about return sets the next quarter and where the next investment goes.

The decision rights and the brief, in full
Decision rights, one owner each
  1. What gets builtThe product owner for that area
  2. What gets declinedThe same person, and it is written down
  3. What gets stoppedWhoever owns the metric it was meant to move
  4. What shipsA named approver, one step, logged separately from the check
  5. A failing evalThe owner of the test set, who can waive it in writing
  6. Money, data models, public interfacesAlways a person. The threshold is the policy
The brief, in full. Five fields, one screen.
  1. 01The problemIn the words of whoever has it, not the words of the fix.
  2. 02The definition of goodThe number, the bar, and the date it gets read.
  3. 03The non-goalsWhat this deliberately does not do, so scope cannot drift silently.
  4. 04The cost budgetWhat one unit of this work is allowed to cost to run.
  5. 05The lineWhat is fixed code and what is a model call, decided before building.
What to count04
Four swaps

Swap what you count and the conversation changes inside a week.

Each one needs a definition agreed in advance and somebody willing to read it out when it comes back unflattering.

Spend against landed

Both start at the same index. One more than triples. The other does not move.

Illustrative
0100200300

Indexed, Q1 = 100

Spend on agent toolingLaunches that changed behaviour
Spend up, landing flat
The shape I keep seeing: spend on agent tooling against the share of launches that changed behaviour.
Instead of launches shippedLaunches that changed a behaviour you named in advance
Instead of cycle timeTime from a customer signal to a decision about it
Instead of AI adoptionCost per unit of work an agent produces, against the margin it protects
Instead of a quarterly reviewA weekly read anyone on the team can run without asking you
Next

The point of view is the method. An engagement applies it to you.

From a first read of where you stand to the whole zero-to-one arc.