You still hear the old calendar in the models. Mention a database migration and a large share of assistants reach for project language: multi-week phase, careful planning, several engineers. That estimate is not crazy if a human has to design, babysit, and maintain the work. It is increasingly the wrong unit once a shaped goal can own the outcome overnight.

The lived pattern looks more like this. You give an agent a North-Star outcome — not a script of keystrokes — plus how it will know it is done, plus a few quality gates. You leave. You get a coffee. Sometimes you come back to done, and the done is real. Sometimes the path is aesthetically offensive and still correct. The economic question that matters is no longer whether a senior engineer would have been proud of the intermediate. It is whether the world loop closed.

All that cost me is the intent of defining it.

That line is not claiming intent is free. It is claiming the relative prices inverted. Machine time, shell time, and iteration patience became abundant. The scarce collaborative act is shaping an owned outcome tightly enough that autonomous labour can finish it without you sitting in the loop as a meat-based debugger. Frontier systems already sustain multi-hour tool-using runs on open-ended problems; the useful horizon has been expanding on a timescale of months, not decades.¹ Anthropic’s framing of agents emphasises multi-turn tool direction on unpredictable-step work — not single-shot answers.²

Flip the budgeting question

The old budgeting question was: how much work will this take? The new question, for operators who already run multi-hour autonomous goals, is sharper: can I form a sufficiently good goal? If yes, execution capacity is no longer the binding constraint. If no, more agents only multiply wrong-altitude motion.

We have argued elsewhere that AI behaves like an intention compiler: it multiplies what you specify and mirrors what you leave vague.³ This piece is about the next economic fact: under abundant machine time, the scarce input to that compiler is goal formation itself. How you write orienting purpose in detail is its own craft;⁴ the point here is the cost structure, not the template.

Reproductive rate greater than one

There is a quiet expectation buried in productivity culture: if you work hard enough, the list should empty. Operators running real agent fleets keep violating that expectation in the most flattering way possible. They complete more. The list grows faster.

one thought
   ↓
one shaped goal
   ↓
hours of autonomous work
   ↓
new capability / new data / new discovery
   ↓
three new thoughts

Your system has a reproductive rate greater than one. Each completed goal can create more than one new plausible goal. Of course the backlog grows. There may be no natural point where you honestly say you have run out of things for the agents to do. The agents are increasing the reachable idea surface.

Previously, implementation drag throttled cognition. You had an idea, hit a technical problem, spent two days coding, and only then resumed thinking at full bandwidth. Remove the drag and the process becomes catalytic. Even with the same memory, capture and formation failure rates can rise simply because more perishable objects arrive per hour. External research keeps rediscovering a cousin of this pressure: AI tools can raise output while also intensifying workload rather than simply deleting effort from the human side of the ledger.⁵

If R is greater than one, grinding without formation quality is a treadmill with a motor. The only throttle that matters is goal quality — which goals you promote, which shapes you refuse to collapse early, which higher-order purposes you recognise before you spend a week of agent labour on the wrong altitude.

The scarcity-inversion ledger

After a week of multi-hour goals, ask what actually ran out. Not disk. Not shells. Clear intent.

For most of software history, human intent was relatively cheap to state and expensive to execute. Now a sentence can unlock hours of autonomous labour — if and only if the sentence is really a goal. Intent stopped being the free preamble. It became the binding input.

volatile thoughts
      ↓
intent archaeology (longitudinal inference)
      ↓
big-block reasoning
      ↓
GOAL FORMATION          ← scarce human/AI collaboration
      ↓
abundant execution

Industry language already frets about uncontrolled proliferation of agents — sprawl without commensurate architecture for how work is chosen and steered.⁶ Sprawl is what abundant execution looks like when formation is weak. GitHub’s framing of AI-era development keeps drifting toward a blunt slogan: intent is becoming the source of truth, with code as a rendering of that intent.⁷ If that is even directionally right, organisations still budgeting as if the scarce resource were coder-hours are optimising last decade’s ledger.

POSTGRES! LET’S GO — premature specificity

Say “old mail database” into an execution agent. Watch what happens next:

POSTGRES!
SCHEMA!
ETL!
JSONL!
LET’S GO!

The comedy is load-bearing. The agent is not stupid. It collapses ambiguity into actionable technical work. Motion feels like progress. And sometimes — once you have already chosen the field — that collapse is exactly what you paid for.

The failure is timing. You had a shape, not a task. Call the failure mode what it is: premature specificity — collapsing fuzzy intent into implementation detail before the purpose is stable. You still need precision about success and failure. What you must not do early is precision about the wrong altitude: table names, vector indexes, YAML fields, tool shopping.

DeepMind’s warning about specification gaming is a cousin: systems that satisfy the literal objective without the intended outcome when the objective is poorly shaped.⁸ Premature specificity often creates that risk in reverse — a brilliantly executed pipeline for a goal you never actually chose.

Want the rabbit for execution. Do not hire it to decide which field you are standing in.

Management vocabulary is still one level too low. Assign, monitor, review assumes Task A already exists. Your real problem often starts in the mush where thought is still a wooden block with no label on the underside. Put the blocks on the table — OLD EMAIL, DEV PROJECTS, CLAUDE HISTORY, EBOOKS / IP — and sit there. Do not implement. Do not open a migration. This is a cousin of progressive resolution: stabilize coarse structure before high-resolution detail.⁹

Intent archaeology: Monday to Thursday

You do not always know what you are building while you are building it. That watching-from-above is intent archaeology: look across several days of exhaust and name the buried thing.

A low-level agent summarises that week as five ingestion sources. The natural next move is five pipelines. That summary is not false. It is incomplete in a way that costs the whole game.

You have progressively expanded from indexing finished intellectual output to reconstructing the history of your cognition and work.

Once that sentence is on the table, the week reorganises. The ebooks were finished output. The dev folders were the making. The conversations were the deliberation. The mail was residue around the work. The ancient archive was the deepest layer of the same history — not a random side quest. Possible higher-order purpose: allow future agents to understand not only what is known, but how ideas and decisions emerged over time.

The recognition moment is the method. When the steward puts the block down and you hear yourself say some version of oh — yes, formation just happened. Only then do you earn the rabbit. Archaeology is reconstruction from remains: a large share of real intent is never stated as a clean Monday sentence. It is enacted across reaches. Capture freezes a volatile thought; archaeology asks what the sequence of thoughts is becoming. Capture and dispatch mechanics matter and are covered elsewhere;¹⁰ they are not this article’s product.

The intent steward

What you need is not better management of existing tickets. It is an intent steward: field-chooser, archaeologist, altitude guard — not another rabbit.

• Not a manager — managers assign Task A/B/C that already exist.
• Not a rabbit — execution agents dig; you want them after the field is chosen.
• Is a steward — holds big blocks, runs longitudinal inference, refuses premature schema talk, proposes higher-order goals for confirmation.

Variants of the same doctrine: a weekly human partner session; a personal steward agent altitude-locked by design; a team ritual that asks “what is the giant block?” before rabbits are released. Success metric: correct altitude and confirmed big blocks — not tickets closed. If your “steward” pivots from blocks to Postgres mid-sentence, it is an unmarked rabbit with better manners.

Hours, not weeks

The steward needs an empirical model of your throughput — not a generic PM textbook.

Not everything becomes cheap. What remains expensive is exactly this article’s subject: forming goals, naming purposes, holding altitude. Do not reject a block because it “sounds like three weeks” if your system has already turned that class of work into goal slots. And do not accept it without formation — releasing a rabbit into an unnamed field.

The doctrine in one breath

When execution, iteration, and machine time become abundant, the binding constraint is the human’s ability to form a good goal. The highest-leverage role is an intent steward that infers, longitudinally, what you are really building and keeps you at big-coloured-block altitude instead of collapsing into specifics. The failure mode is premature specificity. The method is intent archaeology. The economics are the scarcity-inversion ledger and a backlog with R > 1.

None of that deletes excellent rabbits. It sequences them. Form the goal. Confirm the block. Then let the dig be absurd, patient, and unattended. Leave the SQL alone until the coloured block is on the table — and once it is, stop romanticising the intermediate file format. Ask whether the world closed.