← All Posts

Unit of work

The contract that replaces the prompt: a job with a definition of done, a budget, and a single owner, living in a workspace.

June 5, 2026·4 min read

Phase 1 said businesses hire agents and pay for outcomes. This page defines the thing they actually hand over: not a prompt, a job.

A prompt is a single instruction with no memory, no files, no tools, no budget, and no record once the reply scrolls away. Whatever an agent does in response is unowned and unverifiable, and every serious deficiency of prompt-driven work traces back to that statelessness. A unit of work is the opposite: a job that is owned from start to finish, executed, checked, and paid for. None of that fits in an instruction; it fits in a workspace.

A prompt
  • One instruction
  • No memory or state
  • No tools or files
  • Nothing to verify or bill
A workspace
  • Owns the whole outcome
  • Memory and files persist
  • Tools to reach your systems
  • Budget, state, and record built in

What the job needs to live

Runtime
A place to actually act, not just answer.
Memory
Context that carries across the whole job.
Files
The artifacts the work produces and edits.
Tools
Access to the systems where the work lives.
Budget
A bound on what reaching the outcome is worth.
Record
A trail to verify what was done and what it cost.

The lifecycle

01
Define
Name the outcome and the end state you want.
02
Budget
Set what that outcome is worth.
03
Execute
The agent works inside the workspace.
04
Verify
Compare the state, before against after.
05
Settle
Budget against the AI spend on your model.

What the contract buys you

Definable
The end state is written down before work starts.
Verifiable
Done is a state comparison, not a self-report.
Priceable
Budget vs. spend, per outcome, not per token.
Accountable
Exactly one owner per unit.

A definition of done forces the requester to say what correct looks like before execution, which is where most delegation failures are created. Verification by state comparison is robust in a way that reading the agent's self-report is not. A budget makes the gap between value and spend a clean per-outcome signal. And one owner per unit means that when something ships broken there is exactly one place to look, the property that lets a business extend trust to a non-human worker at all.

There is a subtler reason the unit is built this way, borrowed from optimization theory: any proxy placed under pressure gets gamed (Goodhart's law), and activity metrics (messages sent, tokens generated, hours logged) are proxies an agent optimizes effortlessly. A verified state change is the hardest-to-game measure available, because it is the outcome itself.

The loop a business already runs

01
Intent
What the business wants: a definition of done plus a budget.
02
Action
Execution inside the workspace, by a person or an agent.
03
Impact
The verified state change, settled against the budget.

Businesses do not think in models, agents, or tokens, and should not have to: nobody prices an accountant by spreadsheet keystrokes. The loop they run is intent, action, impact, and the unit of work maps onto it one-to-one. The model, the agent, and the tokens disappear inside the action step as swappable implementation details, which is what lets an agent's unit of work sit in the same queue, carry the same kind of budget, and be judged by the same check as work done by people.

The contract defines and verifies the job. What it does not do, by itself, is give the delivered work a unit of account. Next: the quirq, the unit minted when this lifecycle completes.