Resources / agentic system design 01

Can this proof of concept reach production?

Most agentic proofs of concept get judged after the demo goes well, which is the one moment nobody is able to judge them. Twelve questions, scored before you build. Each one says what earns a 0, a 1 and a 2, so two people scoring the same thing land in the same place.

12 questions · 4 sections · max 24 · one hard gate

0/24
Incomplete 12 questions unscored. A blank is a 0.
0/12 scored
00 How to run it

The rules that decide whether this is useful or theatre.

  1. Score what is true today, not what is planned. "We'll build that in the pilot" is a 0. The screen measures readiness, not intent — this is the single rule that decides whether the exercise is useful or theatre.
  2. Everyone scores privately, then reveals. Read the question aloud, each person commits to a number, then compare. Discussing first collapses the room onto whatever the loudest person thinks.
  3. A two-point spread is the finding. Where scores disagree by two, the team does not share a picture of the system. Resolve that before moving on; it is worth more than the total.
  4. A blank is a 0. If nobody in the room can answer, nobody knows, and not knowing is the condition being measured.
  5. Re-score after narrowing, not after arguing. The legitimate way to move a score is to change the scope of the POC.
A

The work

Whether the task is shaped for an agent at all. These three fail quietly — nobody notices until the evaluation has nothing to measure against.

01 This task happens in substantially the same shape hundreds of times a month.

Repetition is what amortises the build. High-variance, low-volume work never earns back the evaluation and review cost.

02 You can write the complete list of actions the agent may take, and it is short.

If the list is open-ended, so is the blast radius, and no reviewer can approve it. Count the actions before you count the features.

03 There is a record of the correct outcome somewhere you can test against.

Closed tickets, ledger entries, prior decisions. No recorded ground truth means no evaluation, and no evaluation means the demo stays the only evidence you will ever have.

B

The blast radius

Hard gate

Whether you survive being wrong. A zero on any row here is a hard stop regardless of the total — this is the section that ends careers, not roadmaps.

04 The worst single action it can take is reversible, or gated behind a human.

Refunds, sends, deletes, commitments. The classic failure is a tool call that times out rather than fails, leaving the system unsure whether the action happened — and a sensible retry that does it twice.

05 A wrong answer is caught by someone before it reaches a customer, a regulator or a ledger.

Silent wrongness is the expensive kind. If nothing downstream catches it, the evaluation is your only defence and it will not be enough in week one.

06 A named person can stop it within minutes, and you can reconstruct afterwards why it acted.

A kill switch nobody owns is not a kill switch. A trace you cannot replay is not an audit trail.

C

The evidence

Whether you will be able to prove it worked. Decided before the build or not at all — you cannot retrofit a baseline.

07 You know what this task costs today in time, money and error rate.

Without a measured baseline, “better” is an opinion, and the pilot review becomes an argument about anecdotes.

08 You can produce the hard cases today: missing information, conflicting history, two near-identical requests at once.

These are the cases a demo never contains and production contains hourly. If you cannot assemble them now, you will not assemble them under launch pressure.

09 You can test what it does when a tool returns neither success nor failure.

Timeouts, partial writes, ambiguous responses. Most agent incidents live here, and almost no POC tests it.

D

The ownership and the economics

Whether it survives contact with the organisation. Pilot purgatory is an ownership problem far more often than a tooling one.

10 One named business owner — not a committee — has a number that improves if this works.

If no one's metric moves, no one defends the budget at the next planning round, and the pilot quietly ages out.

11 The data it needs is in systems you can call today.

If a data programme has to land first, that programme is the real project and it is six months long. Say so out loud now.

12 Your cost estimate includes tool calls, retries and human review minutes — not just the model bill.

At a hundred times the volume the model is usually the smallest line. Review labour is usually the largest, and it is the one that scales linearly.

05 The cut-lines

Agree these before you score, not after.

A total that lands one point above a line is a narrow, not a yes.

Any 0 in B
Hard stop, whatever the total
You cannot buy your way out of blast radius with a high score elsewhere. Fix the reversibility, the review step or the stop control first, or pick a different POC.
20–24
Pilot candidate
Write the boundary down — who can use it, what it may do, which actions need review, who can stop it — and run a limited trial against it.
14–19
Narrow it, then re-score
Usually the action list is too wide. Cut permitted actions until the low-scoring rows move. A smaller agent that ships beats a broad one that stalls.
0–13
A demo, not a pilot
Not a bad idea — a bad first idea. Build it to learn if you want, but do not put it on a roadmap or in front of a customer.
06 Before you build

Four ways to answer most of these without the agent existing.

In ascending order of cost. Every one of them is cheaper than a pilot that fails in front of a customer.

01

Write the evaluation set before the spec

Ask for fifty real cases from last quarter with known-correct outcomes. Recorded ones, not invented ones. If the team cannot assemble fifty in a week, that isn't a scheduling problem: the task has no ground truth, and an agent you cannot grade is an agent you cannot ship. The failure to produce the set is the verdict.

Cost: one week, no code Answers 03 · 08
02

Put a human behind the curtain

Before building anything, route real requests to a person sitting behind the interface the agent would use, answering under the same rules and with the same tools. You learn the true distribution of requests, which ones have no clean answer, and whether anyone uses the thing at all. You end the fortnight holding a labelled evaluation set you didn't have to imagine. Check the disclosure question with legal first if the flow is customer-facing.

Cost: one person, two weeks Answers 01 · 02 · 10
03

Break the tools on purpose

In staging, make the tool time out rather than fail. Make it return success for an action that never happened. Make it return the wrong thing, confidently. Chaos engineering has been standard for services for twenty years and is almost unheard of for agents — which is odd, because this is precisely where agent incidents live.

Cost: a day of harness Answers 04 · 09
04

Run it in shadow

Put it on live traffic and forbid it from acting. Log what it would have done, compare against what the human actually did, and review only the disagreements — usually a small fraction of cases, and all of the information. A real accuracy number with a blast radius of zero.

Cost: a logging path and reviewer time Answers 05 · 07
Where this comes from

The screen is the easy half. The argument is the other one.

A blank scorecard is genuinely useful and costs nothing to give away. What it cannot do on its own is settle the rows where two people scored two points apart, which is where the real disagreement about the system lives. That part takes a room.

See the cohort  All resources 

Published 13 September 2026 · free to use and to pass on