AGENT-CENTERED DESIGN

Your agents don't have an intelligence problem. They have a decision problem.

A product is the aggregation of every decision made across its lifecycle. An agent workflow is a decision graph with most of the humans removed from the middle. Agent-Centered Design is the practice of building systems around the decisions agents and people make together — and instrumenting them so you can prove it.

JOIN THE LIST READ THE STORY →

THE PROBLEM, MEASURED

Agent operations run on records that nothing reads

Wherever writing is cheap and reading takes effort, systems over-produce. Industry estimates put the majority of telemetry spend on data that is never queried. We measured our own agent operation and found the same shape — then built the instruments to fix it.

~70%

of observability spend goes to log data that is never queried

INDUSTRY ESTIMATE

5 : 1

agents under-report their own friction versus what an independent extractor finds in the record

MEASURED, OUR OPERATION

65%

of wall clock in our first fully delegated sprint was batch-induced waiting, not work

MEASURED, OUR OPERATION

A figure that verifies itself

This is not an illustration. It is our first fully delegated sprint run, rendered from its own event log: four agents, two waves, 863 seconds of wall clock. Toggle to see what the measured working times imply if the scheduler had honoured the dependency graph it computed — and then discarded.

FIRST FULLY DELEGATED SPRINT · RENDERED FROM THE RUN'S EVENT LOG

FRONTEND

BILLING

API

INFRA

0s216s432s648s863s

WORKINGWAITING

WALL CLOCK863s

TIME WAITING561s · 65%

STATUSmeasured

Every bar is computed from the run's event log — agent start, finish, merge. The wave boundary cost 259 seconds of orchestrator turnaround; finished agents sat unmerged for another 303. The batch was one line of scheduling code.

The one-piece view is a target computed from measured working times — not yet a result. The scheduling fix has shipped; the instrumented re-run is how we falsify our own claim.

The method

THE UNIT OF DESIGN IS A DECISION

Not a feature, not a story, not a ticket. Map a workflow into the decisions it actually contains — supplier, input, conversion, output, customer — and you get a graph you can measure, improve and hand to an agent. Most organisations have never drawn theirs.

DISAGREEMENT IS MADE PRODUCTIVE BY MECHANICS, NOT BY MANNERS

Put a panel of agents on a question and they converge on the average of their priors. Adversarial collaboration is the set of mechanics that stops that: a briefing packet instead of a pitch, explicit affordances to refute, an accountable author on every convergence, and dissent kept on the record rather than resolved away.

NO OBSERVABLE IS SELF-REPORTED

Every measurement in the system is computed from the record, never authored by the agent being measured. That single rule is what makes the numbers on this page arguable — and what lets a second reader regenerate any of them.

Every figure carries its status, its source, and what would prove it wrong. Claims are pinned to the record that produced them, and the book refuses to compile when a source no longer says what the prose claims. That is the differentiator, and it is not a marketing posture — it is a build constraint.

THE RESEARCH

What is being built

Three instruments

THE WASTE BOARD

Nine wastes, translated from manufacturing to agent operations, each computed from the record on a weekly cadence. The board names the biggest row; the improvement work starts there.

THE WRITE PATH

The place where a record becomes durable — and therefore the place to enforce that it has an accountable author, a live source, and somewhere to put dissent. Refusals are typed, counted and owned.

THE AGENT PROCESSOR

The scheduler and the seat an agent occupies: what it is handed, what it may write, when it hands back. Most batch waste hides here, in the gap between what the dependency graph says and what the scheduler does.

Nine wastes, nine observables, named owners

OVER-PRODUCTION
Writes that nothing cites within 30 days — records produced for no reader.
OWNER: WRITER
WAITING
Queue age at human gates; agents finished but unmerged.
OWNER: HUMAN
TRANSPORTATION
Handoff items the receiver never cites — context shipped, not used.
OWNER: SENDER
MOTION
Search calls per addressed read — an agent that cannot locate itself reads everything.
OWNER: STORE
EXTRA PROCESSING
Re-deriving a claim that already exists, uncited.
OWNER: WRITE PATH
INVENTORY
Proposed work aging unreviewed, measured against a declared review capacity.
OWNER: REVIEWER
DEFECTS
Typed refusals by reason code; violations accepted that should have been refused.
OWNER: WRITE PATH
NON-UTILISED TALENT
Refusal-eligible states where no agent refused — judgment present, unexercised.
OWNER: HARNESS
SAFETY
Out-of-grant writes; irreversible acts without a reversal pin.
OWNER: WRITE PATH

The Practice Range

The Practice Range is where readers stop reading about the method and operate its instruments. Every lab is an operation of a real instrument, and every lab ends with a measured artifact in the learner's own sandbox. No quizzes, no videos with checkboxes — the transcript is itself a record, which is what makes the certification mean something later.

01
THE PILE
Map a workflow you actually run into decision nodes. While you work, the system measures your own session's lead time, wait and work in progress — then shows you.
MEASURED ARTIFACTYour first decision graph, and the reveal: you were just measured
02
ADDRESS IT
Complete the same retrieval task twice — search-only, then with addresses. Your own motion row moves on the board.
MEASURED ARTIFACTA before-and-after motion ratio
03
STILL TRUE?
Change a source underneath a pinned claim. Watch the compile refuse, and name the pin. Assign the invalidation owner and the cadence.
MEASURED ARTIFACTA broken pin, a refusal, and a named owner
04
THE AVERAGING PANEL
The flagship. Run a small panel handed a pitch and watch it produce a near-zero refute rate and a worthless convergence. Re-run it with a briefing packet and explicit refute affordances. Compare the dissent, measured.
MEASURED ARTIFACTTwo panel transcripts and the refute-rate delta
05
SIGN IT
Author the convergence from the previous lab. The write path refuses it without an accountable author and a non-empty dissent list.
MEASURED ARTIFACTAn authored position with its dissent retained
06
THE BATCH
Dispatch a scripted multi-agent run in waves, then by true dependency graph. Read your own flow diagnostics — the demo above, except it is your run.
MEASURED ARTIFACTYour own batch-wait percentage, twice
07
CAPSTONE
Run the waste board across your whole lab history. Pick your biggest row, and author an improvement with a definition of success, a falsifier and an owner.
MEASURED ARTIFACTAn improvement claim that could survive review

The agent tutor is not a chatbot beside the content. It takes the seat opposite the learner: the panel members in lab four, the reviewer in lab five, the second reader who regenerates — or fails to regenerate — the learner's board in lab seven.

THE BOOK

Agent-Centered Design

ARCHITECTING ADVERSARIAL COLLABORATION IN AN AGENTIC WORLD · CORY NATION

The argument in full: why the unit of design is a decision, what changes when the decider is an agent, and how adversarial collaboration makes panels converge instead of average. Every claim pinned to the record that produced it.

I
THE PREMISE
Every product is a pile of decisions — mappable, improvable, and the one layer nobody maps.
II
THE FRICTIONS
Put an agent in the decider's seat: rebuilt context, search where addressing should be, staleness nobody owns, handoffs that strip reasoning, batches hiding in the write path.
III
THE METHOD
Adversarial collaboration — disagreement made productive by mechanics, convergence with an accountable author, dissent kept on the record.
IV
THE PRACTICE
The method stated end to end, then run on the author's own operation, with the receipts.

PRINT AND EBOOK

Freezes the record at press time. The argument, the figures and the pins as they stood the day it went to press.

THE LIVE EDITION

The same chapters, with the figures computed at read time and the development of the system itself visible as it happens. When a source changes underneath a claim, you see that too.

Nothing is for sale here yet. Join the list and you will hear first — when the book lands, and when early access to the digital edition opens.

THE PATHWAYS

A practice you can be certified in

Training is modular: each module carries its learning objectives, and each certification names the objectives it assesses — so pathways compose from modules rather than duplicating them. Every assessment above the first level judges artifacts from the learner's own sandbox, not multiple-choice recall. That is what makes the credential defensible.

ACD ASSOCIATE
YELLOW BELT
Working knowledge across the domain: the nine wastes, how systems are architected for agent operations, and the value case.
Assessed on: Knowledge assessment
ACD PRACTITIONER
GREEN BELT
Can identify and define architectures for agent-centered optimisation and flow, and has implemented the method on more than one system.
Assessed on: Practice Range labs four through seven, plus a judged capstone artifact
ACD PRODUCT OWNER
BLACK BELTNAMING UNDER REVIEW
Proven application across a system of applications and architectures — owns a portfolio of agent systems and value streams. The first level permitted to train others.
Assessed on: Portfolio evidence, an advanced capstone, and a review board
ACD MASTER
MASTER BLACK BELT
Develops and teaches their own curriculum, and licenses the programme for redistribution.
Assessed on: Teaching practicum and curriculum review

The assessment is an interview

ACD assessments are not multiple-choice tests. The assessor is an agent that interviews the candidate, on terms the candidate chooses, because a certification should measure what a person knows about deploying the material in their own domain — not their skill at taking tests.

FIXED WHAT, FLEXIBLE HOW

The competency requirements are fixed. How each one gets assessed is the assessor's choice, made for this candidate.

THE CANDIDATE PICKS THE TERMS

Chat, voice, or working the environment directly. Comfort is a validity feature, not a concession.

SCAFFOLDING DEPTH IS THE SCORE

On a partial answer the assessor offers guiding context. How much guidance each competency needed — unaided, prompted, guided, or not demonstrated — is the measurement. The bar is competency coverage at a named level of support, never a percentage.

THE INTERVIEWER NEVER SCORES

A separate pass over the transcript and evidence map produces the verdict, so the conversation cannot talk itself into a grade. The record is retained, and a second reader must be able to regenerate the result.

Scaffolding depth as a score is not new — it is dynamic assessment, with a validated lineage going back decades. What is new is applying it at certification grade, to agent-operations practice.

SPONSORSHIP

Sponsor the research

Organisations can fund the development of Agent-Centered Design. Sponsorship buys access and association — early access to research drops and tooling, founder briefings at the upper tiers, and public recognition as a supporter of the work.

SUPPORTER

Recognition on this site as a supporter, early access to research drops and book chapters, and the community around the work.

SPONSOR

The above, plus quarterly founder briefings, early tooling access, and input on the research agenda — input, never control.

FOUNDING RESEARCH PARTNER

The above, plus monthly founder access, the method applied to the sponsor's own agent operations, and first access to certification. Deliberately capped at a small number of organisations.

Money buys access, never conclusions. Findings and the body of knowledge are not sponsor-influenced, ever.
Every sponsor is disclosed, with their tier, prominently.
Concentration is capped, and results are published whether or not they flatter a sponsor.

FOUNDING SPONSORS — INQUIRE.

Tiers and terms are being finalised. This is a conversation, not a checkout — tell us you are interested and we will take it from there.

EARLY ACCESS

Join the list

The system is being built with its own method — every decision recorded, every dissent kept. Watch it happen, or put the instruments on your own operation.

One list. Sign up once and you are in the immediate funnel — and the same address becomes your account when the digital edition opens. No password to invent today.

You will hear when the book lands, when early access to the digital edition and the Practice Range opens, and when the research moves. Nothing else.