THE AGENT FORGE WAY · VERSION 1
How we build, in fourteen rules.
Everyone can buy the same models, so the models are not what you are paying for. This page is. It is a way of operating: how a process is sorted, what a model is allowed to decide, what has to pass before anything goes live, and what you can reverse without us. An engineer joining one of our sprints is handed it on day one. A client doing due diligence can read it before signing. And every claim on this site points back to a rule on this page, so you can hold us to it.
Version 1 · 6 October 2026 · revised as each sprint teaches us something, with the change noted below
Who this is for
An engineer joining a sprint
Read all of it. The rules are the job; the section near the end is what you are handed and what you own.
A client, before signing
Rules 01 to 14 and the due-diligence section. Ask us about any line you would like in the contract; most of them already are.
Anyone checking a claim
The last table maps every sentence on the home page to the rule that backs it.
The fourteen rules
Two of twelve steps use a model. A person owns the exceptions. Everything is logged. The rules below are how that stays true under deadline.
01
Baseline before build
We measure the process before we touch it: cycle time, cost per case, straight-through rate and hours a month. The numbers are written into the sprint contract and re-measured at day 90.
Enforced by: Contract
On a sprint
You get the baseline sheet before you get system access. If a number on it is wrong, say so in week one, not at the readout.
02
One workflow at a time
A sprint builds one workflow end to end. The second workflow is written down, priced, and not started. An owner who wants three things at once gets the one with the biggest measured loop. The first sprint also builds the spine the next ones run on: one intake, one approval route, one log, one place a model is chosen. That is why the second workflow is cheaper and faster than the first.
Enforced by: Scope
On a sprint
Anything the client asks for outside the sprint scope goes to Rui the same day. You do not say yes or no on site.
03
Delete, code, agent, human, in that order
Every step in the mapped process is sorted: delete it, handle it with a deterministic rule, give it to a model, or leave it with a person. A model is the third choice, not the first.
Enforced by: Method
On a sprint
The sort is done at the assessment and is in your brief. If you find a step that should move, propose it with the reason; do not move it quietly. And before a step goes to a model, check whether the answer is already known: a decision the agent has made before on the same input is looked up, not asked again. Remembering is code, and it is cheaper than the smallest model.
04
The model only where judgement is needed
Models read documents and classify them, the things a rule cannot do. Rules match and check tolerances. The agent then updates the system of record directly, with the permissions the client sets, and every update is logged. A model never decides on its own whether a number is acceptable.
Enforced by: Code
On a sprint
If a prompt is doing arithmetic or a tolerance check, it is in the wrong place. Move it to a rule.
05
The eval gates the deploy
At least fifty golden documents from the client's own history, chosen with the person who does the work today to include the edge cases and the high-stakes ones, not just the typical ones. The agent does not go live until it passes all of them, the pass count is the first line of every run log, and the set is re-run on every change after go-live, not only before it.
Enforced by: Code
On a sprint
You build the golden set with the client's clerk in week one. Edge cases they remember are worth more than the ones you invent.
06
Shadow, then exceptions, then live
The agent first runs beside the person and posts nothing; every difference is reviewed. Then it posts the matches and routes the exceptions. Then it runs. The client moves the dial at each stage, not us.
Enforced by: Product
On a sprint
Shadow mode is four weeks, not a day. The daily diff is a deliverable. Do not ask to skip it because the eval passed.
07
A named person owns every exception
Anything outside tolerance goes to a named approver, in the channel they already use, with the reason. Approval thresholds are the client's policy, unchanged by us.
Enforced by: Product
On a sprint
The approver's name and the thresholds are in the brief. You do not change either. If the client wants to, they tell Rui and it goes in the change log.
08
Every action logged and reversible
What was read, what was decided, what was posted, by which step, at what time. A wrong post can be found and reversed in one place, by the client, without us.
Enforced by: Code
On a sprint
If you cannot show the client how to reverse a posting from the log, the sprint is not done.
09
Inside the system of record; nothing new to learn
We build inside the ERP, CRM and inbox the client already runs. Approvals land where approvals already land. A client never logs into anything of ours.
Enforced by: Product
On a sprint
No dashboards, no portals, no new logins. If a screen is needed, it is a report inside their system.
10
The cheapest model that passes
Each step that needs a model is benchmarked on its own against several, small local models included, and the one that passes the eval at the lowest cost runs that step. Most steps need no model at all. Every agent has a hard ceiling on calls, tokens and run time that it cannot exceed. The run cost on the proposal is the real one, and it sits on the readout next to the hours returned.
Enforced by: Code
On a sprint
Benchmark per step, not per workflow; the step that reads a messy PDF and the step that picks a category rarely want the same model. Benchmark results go in the sprint folder. A model is swapped only when the new one passes the full golden set again.
11
Your data stays in your environment
Model calls go through the client's own cloud provider, in a Canadian region where one is available, under zero-retention terms. Nothing is used to train anything. Nothing leaves the client's environment without their sign-off.
Enforced by: Contract
On a sprint
No client documents on your own machine, in your own accounts, or in any model you reach outside the client's provider. This is the rule that ends an engagement if broken.
12
Re-measured at day 90, in the client's numbers
The same four numbers as the baseline, measured the same way, on a date fixed in the contract. The readout is one page. If the numbers did not move, the readout says so.
Enforced by: Contract
On a sprint
The re-measure date is booked before the sprint ends. The numbers come from the client's system, not from our logs.
13
Handover with a runbook; staying on is the client's choice
Every sprint ends with a runbook the client's own people can run from, and with the controls to go with it: the business rules and thresholds live outside the code so your process owner can change one without an engineer, and one screen shows what is running, pauses it, and rolls a post back. Run and maintain, where we monitor, re-test and swap models, is offered, never required.
Enforced by: Contract
On a sprint
You draft the runbook from week eight, and the client's person walks through it with you before day 90. If they cannot follow it, it is not finished.
14
“Not worth automating” is a real answer
An assessment that ends in a recommendation not to build is a complete assessment. We say it on first calls when it is true. A process that is only slow because of one policy gets the policy fix, not an agent.
Enforced by: Practice
On a sprint
If the loop you are automating turns out to be a policy problem, stop and say so. Nobody is paid to build the wrong thing well.
The ninety days
Dates are fixed in the sprint contract on day one. The shape is the same on every sprint; the client's dial decides the pace of the last third. On a repeat-stack sprint, where we have already built the same workflow on the same system, the build shrinks to two weeks and exceptions and live to one week each. Shadow stays four weeks: eight weeks in all.
Before day 1
Assessment
The map, the sort, the baseline and the priced scope. Rules 01, 02 and 03 are done here, by Rui, before an engineer is assigned.
Weeks 1–3
Build to the eval
Golden set assembled with the client's clerk. Rules for the deterministic steps, prompts for the two model steps, the exception route and the log. Done when all fifty pass.
Weeks 4–7
Shadow
The agent runs beside the person and posts nothing. Every difference is reviewed daily. The threshold for moving on is agreed with the client in writing.
Weeks 8–10
Exceptions
Matches post. Exceptions go to the named approver with the reason. The runbook is drafted. The client can move the dial back at any time.
Weeks 11–12
Live
The agent runs. The client's person walks through the runbook. The re-measure date is confirmed.
Day 90
Readout
One page: the baseline, the same numbers now, the run cost beside the hours returned, what was not measured, and the second workflow priced and not sold.
If you are joining a sprint
You are a forward-deployed engineer: you build in the client's office, inside their systems, under their policies. Rui scopes, gates and reads out. You make it run.
What you are handed on day one
- The assessment brief: the process map, the sort, the baseline sheet, the named approver and the thresholds.
- The reference agent for the client's stack, with its eval harness and the golden-set template.
- The shadow-mode protocol and the daily diff format.
- The exception-routing pattern for Slack or email.
- The runbook template and the handover checklist.
- Access, provisioned by the client in their own accounts, read-only until shadow mode.
What done means
- All fifty golden documents pass, and the pass count is in the run log.
- Four weeks of shadow mode with the daily diff below the agreed threshold.
- Exceptions reach the named approver with the reason, in their channel.
- Live, with the dial in the client's hands.
- The runbook accepted by the client's person, in writing.
- The re-measure date booked.
When to stop and call Rui
- The client asks for anything outside the sprint scope.
- A threshold, an approver or a policy would have to change for the build to work.
- You find data you were not told about, or a system you were not given access to.
- A golden document fails and the fix would change behaviour on the others.
- Anything that would take client data outside the client's environment.
Who owns what
| Rui | You |
|---|---|
| Scope, the sort and the baseline | The build: rules, prompts, routing, log |
| The eval gate: deciding it has passed | The golden set, assembled with the client's clerk |
| Every conversation about numbers, price or scope | Day-to-day work with the people who run the process |
| The day-90 readout | The runbook, and the walkthrough with the client's person |
| Model choice, from your benchmark | The benchmark |
Engineers are subcontracted and non-exclusive, paid a fixed fee per sprint on the same four milestones the client pays on, and bound by the client's NDA and these rules. Rule 11 is the one that ends an engagement if broken.
If you are checking us out
The questions a controller, an IT lead or an acquirer's advisor asks, with the rule that answers each. Any line here can go in the contract.
- Who gets access to our systems, and how?
- Rui and any named engineer on the sprint, through accounts you create and own, with the least access the step needs. Read-only until shadow mode. Revoked at handover unless you keep us on.
- Rule 06, 09
- Where does our data go?
- It stays in your environment. Model calls go through your own cloud provider, in a Canadian region where one is available, under zero-retention terms. Nothing trains anything.
- Rule 11
- What if the agent posts something wrong?
- Every action is logged with what was read, decided and posted. A wrong post is found and reversed from the log, by your people, without us.
- Rule 08
- How do we know it works before it touches production?
- It passes fifty documents from your own history, then runs for four weeks beside your person posting nothing, and you review every difference. You decide when it moves on.
- Rule 05, 06
- Who approves exceptions?
- The person who approves them today, at the thresholds in your policy, in the channel they already use. We do not change your policy.
- Rule 07
- Why not let the model handle the numbers?
- Because it gets them wrong often enough to matter. On real finance work the best model gets 56% of tasks fully right (Vals AI, Finance Agent v2, October 2026); across 19 models on 101 accounting tasks none clears 70% (DualEntry Labs, 2026); Microsoft's launch guidance for its Excel COPILOT function advised against using it for calculations or compliance work, and the function was withdrawn in September 2026. Models read and classify. Rules match, check and post.
- Rule 04
- Which models, and can they change?
- The cheapest that passes your eval, benchmarked and recorded. A model is swapped only when the new one passes the full golden set again, and you are told.
- Rule 10
- What do we own at the end?
- Everything built for you: the rules, prompts, golden set, log and runbook, inside your systems. There is no licence and nothing of ours you have to keep paying for.
- Rule 09, 13
- What about PIPEDA and Law 25?
- The build is designed to whichever applies. Personal information is handled under your policies, and the log is the record.
- Rule 11
- Who are the people?
- Rui Luis scopes, gates and reads out every engagement. Engineers are named in the contract, subcontracted, and bound by your NDA and these rules.
- Rule Engineer section
- What do you not do?
- No platform, no licence, no pilot, no retainer before a measured sprint, no “AI strategy” deck. If the honest answer is that a process is not worth automating, that is the assessment.
- Rule 14
The words we use, and the ones we do not
Says
- Process, step, loop, exception. The words from the map, which you can check against your own Friday.
- Baseline, straight-through, cost per case, days to close. The words from the controller's statement, measurable without us.
- Inside your ERP. In production. Logged. The words that separate a build from a pilot.
- Not worth automating. Said out loud when it is true.
Never says
- AI transformation, digital, agentic, copilot. The vendor's vocabulary; each one puts us beside something free.
- Any model's name in a promise. The answer changes quarterly; the eval does not.
- Pilot, proof of concept. The shadow month inside a paid sprint is never called a pilot.
- Team, where it would mean more than it does. The people on your engagement are named.
Others who say this, with no reason to flatter us
None of these people know us. Each says, from their own work, what one of the rules says.
| They say | The rule |
|---|---|
| “A blueprint tells the system when to involve the LLM at all.” Six months after starting fully agentic, 65% of his workflow nodes were plain code and 29% of workflows needed no model. — Tomasz Tunguz, Theory Ventures, February 2026 | 03. Delete, code, agent, human, in that order. |
| “You're burning tokens to solve a simple problem… hitting a thumbtack with a sledgehammer.” Narrow local models near the data beat a hosted frontier model at twenty to thirty times the cost. — David Linthicum, thirty years of enterprise architecture, October 2026 | 10, 11, 14. |
| The harness — the proprietary data, workflows and context that let an enterprise switch between models — is the layer worth owning. — Chamath Palihapitiya, on where value sits as models become interchangeable, 2026 | 10. Swap the model, keep the build. |
| Run the procedure as code and call the model only at fixed points. With the same model (Claude Sonnet 4) on the TravelPlanner benchmark, that design passed 35.6% of trips against 18.0% for the best model-led method, with 11 commonsense violations against 275. — Qiu, Ye et al., Blueprint First, Model Second, 2025–26 | 03, 04. The model only where judgement is needed. |
| “A year ago, running a couple of agents took us about 30 minutes a day. This week it was eight hours a day.” Agents that make decisions stop failing loudly and hand back work that looks finished. — Jason Lemkin, SaaStr, on running a company on three people and 21 agents, October 2026 | 06, 07. Someone owns every exception after day 90, not only before it. |
| Agree the business outcome, the baseline, the evidence expected at the next review and who owns delivery, before the AI discussion starts. — BCG, CEO's guide to closing the AI knowledge gap with your board, 2026 | 01, 12. The readout is that document. |
Every claim on this site, and the rule behind it
| The site says | The rule |
|---|---|
| “AI that reaches production” | 05, 06, 12. Nothing is called done until it runs live and is re-measured. |
| “Built inside your systems” | 09. Nothing new to learn, no login of ours. |
| “Measured against a baseline” | 01, 12. The same four numbers, before and at day 90. |
| “No platform to sell” | 09, 13. There is no licence; you own the build. |
| “Ships in 90 days” | The ninety days above, with dates in the contract. |
| “First agent live in 90 days, every one measured in dollars” | 01, 12, and the ninety days above. |
| “Every step gets sorted before anything gets built” | 03. Delete, code, model, person, in that order. |
| “Works with the systems you already have” | 09. If it has an API, an export or an inbox, the agent works through it. |
| “Approvals arrive in Slack or email” | 07, 09. |
| “Whichever model passes the evaluation at the lowest cost” | 10. |
| “Models run through your own cloud provider, under zero-retention terms” | 11. |
| “Every agent action is logged so it can be reviewed” | 08. |
| “Including “don’t build this”” | 14. |
| “Who runs it after day 90 is your choice” | 13. |
| “Swap the model, keep the build” | 10. Benchmarked against several; a swap re-runs the golden set, never the build. |
| “Both halves in one room: an operator, engineers who ship inside your ERP, a harness with the eval built in” | 03, 05, 09, and “Who owns what” in the engineer section. |
| “It is built so that it cannot fail loudly” | 05, 06, 08, 14. Fifty documents, four weeks posting nothing, every action reversible, and no-build as an answer. |
| “A model never decides whether a number is acceptable” | 04. |
| “The second workflow is cheaper than the first because it runs on the same spine” | 02. |
| “Your process owner can change a rule without an engineer” | 13. |
| “An agent can only do what you have authorized” | 04, 05, 06, 07, 08, 10, 13. Your permissions, fifty cases, four weeks posting nothing, a named approver, every action reversible, a hard ceiling, a pause. |
Changes to this page
- 14 Oct 2026 · Prices and company-size tiers removed from the site: every engagement is quoted after a first conversation. A home-page answer on what stops an agent doing anything you have not authorized, mapped to rules 04 to 08, 10 and 13 in the table above.
- 12 Oct 2026 · The assessment is a fixed fee again, credited in full against a sprint. Rule 14 still holds: an assessment that ends in a recommendation not to build is complete. Three claims from the new home page added to the table.
- 9 Oct 2026 · The one-workflow assessment is free; rule 14 says a no-build answer is complete, free or paid. Rule 05 now says how the golden set is chosen and that it re-runs after go-live. Rule 02 names the spine. Rule 10 benchmarks per step, includes local models, and caps every agent's spend. Rule 13 adds the process owner's controls. Added the outside voices and five new claims to the table.
- 8 Oct 2026 · rule 04 · Rule 04 reworded: agents update the system of record directly, with the client's permissions and every update logged. A model still never decides on its own whether a number is acceptable.
- 8 Oct 2026 · The home page now shows where the method applies beyond invoices, how a build sits on your systems of record and updates them (rule 09), and how an agent earns autonomy inside a sprint (rule 06).
- 7 Oct 2026 · The home page now says why not a consultancy or a lab. Its two claims were added to the table above.
- 7 Oct 2026 · Added the eight-week repeat-stack sprint, for a workflow we have already built on the same system. Shadow mode stays four weeks on every sprint.
- 6 Oct 2026 · Version 1, written before the first subcontracted sprint. Rules 05, 06 and 13 will be revised with what the first sprint teaches.
Hold us to it.
A thirty-minute call. You leave knowing which workflows are worth handing to agents first and what the loops cost today, whether or not we work together.
30 minutes · no deck · no cost. Bring the work that piles up.