AGENTS

What runs.Apache 2.0.

The only section that does not end in more reading. It is free because it is not the product — it is the proof of the thesis: if the core of the business runs as software, the honest way to argue it is to hand over software that runs.

An organization does not move by departments. It moves by capabilities, and every one of them shares the same bottleneck: hundreds of documents nobody has read together. Minutes, schedules, reconciliations, contracts. Each one was read by someone, once. Nobody has crossed them.

A market of agents, organized by capability. Project governance is built, with three agents (Vera, Samuel and Alba) and a server; CFO and CLO are declared and not yet built. The figures on this page are third-party and carry their source, year and sample. Our own come from the tests and are on each family’s page, with what they measure and what they do not.

Work tools help produce one more document. The problem was never a missing template.

THE RULES

The previous section defined the contract. This one is the code that keeps it.

Read alongside the six conditions →

This is not a collection of loose assistants. Every agent is built on the same five rules, and that is the reason its output can be put in front of a committee.

  1. 01

    Declaration and evidence stay separate

    What someone asserts goes in one column; what the documents support goes in another. They are never merged, and the gap between the two is usually the most valuable finding.

  2. 02

    Nothing is overwritten

    A record keeps its history. When a new version silently replaces the previous one, what disappears is exactly what needed to be seen.

  3. 03

    The model extracts, code computes

    Reading a date is reading. Subtracting, projecting and adding is arithmetic, and arithmetic lives in code, where it is deterministic and can be tested. No number in a report comes from a model estimate.

  4. 04

    Every data point carries its citation

    Or declares that it is absent. "It is not stated anywhere" is a valid and expected finding.

  5. 05

    They know when to stay quiet

    They run on their own and speak only when something crosses a threshold. An agent that reports every week whether or not there is news gets ignored within a month.

WHY THE THIRD RULE IS NOT A TECHNICAL DETAIL

94%

of operational spreadsheets audited in a field study contained errors. An agent that estimates numbers instead of computing them adds one more layer to that same problem.

Powell, Baker & Lawson, Journal of Organizational and End User Computing, 2009 — 50 audited spreadsheets, 270,722 formulas.

The study is from 2009. It is cited as the best available evidence, not as a current measurement.

THE CAPABILITIES

Three capabilities. One of them runs today.

Listed with their real status, because a capability that is declared and not built is still a commitment. What each one costs an organization today is measured by third parties, with the source, the year and the sample stated.

PMO
Built · Vera, Samuel, Alba and Rostrum

Portfolio and project governance

The function that answers for the set of projects: what is running, against what plan, on what budget and with what risks.

Meet Vera, Samuel and Alba →

73%

average cost overrun on IT projects. 18% overrun by more than half.

Flyvbjerg et al., Project Management Journal, 2026 — 11,011 projects, 126 countries.

1/3

of projects have no baseline at all, and 22% are still planned in a spreadsheet.

Wellingtone, State of Project Management, 2026 — more than 150 organisations.

A third with no baseline means a third of projects have nothing to be measured against. That is not a tooling problem — nobody went back to look.

CFO
Declared capability

Close, control and financial reporting

The function that closes the books and answers for what they say: close and consolidation, reconciliations, accounting against tax basis, supervisory reporting and audit preparation.

44%

closed the quarter within six working days in 2023 — down from 49% in 2019.

Ventana Research / ISG, Smart Financial Close research, 2023.

45%

of SOX controls are still manual, while program hours rose 32% in two years.

KPMG, SOX Survey 2025 — roughly 150 professionals.

The uncomfortable figure is the first one: the close did not improve in four years, despite everything that was spent on technology.

CLO
Declared capability

Contracts, compliance and legal risk

The function that answers for what the company signed: contract review, the repository of what was signed, obligations and expiries, regulatory compliance and disputes.

8.6%

of contract value eroded on average by poor contract management. Best a little over 3%, worst more than 20%.

World Commerce & Contracting with Deloitte, 2023 — 1,236 organizations.

90%

face challenges locating their own contracts. A large organization manages about 19,000 a year.

EY with Harvard Law School, 2021 — 1,000 professionals, 22 countries.

A company that cannot find its own contracts cannot know what it signed, what expires, or what it committed to.

The list is not closed. A capability enters when someone who practises it wants to build its agent.

WHAT CAN BE CHECKED

The arithmetic is testable. Run it yourself.

The figures from real runs (how many documents, how long it took, how many findings, and which ones nobody had seen) go up when they exist, with the procedure to reproduce them. What can already be run is the verification of the code.

python3 scripts/verificar.py

everything the repository verifies, for the three agents at once

python3 plugins/criterio-portfolio/scripts/portafolio.py selftest

Vera’s arithmetic against known results

Figures that have not been measured yet are not invented.

The numbers on this page are third-party and carry their source, their year and their sample. Ours come from the repository’s tests and are on each family’s page, with what they measure and what they do not. While a capability has no real runs, its page says so.

THE ECONOMICS

The code is free. Running it is not.

Technical performance is one variable. An AI case creates value when the economic benefit of the work or decision it changes exceeds the full cost of operating, supervising and correcting the system. A working AI system is not necessarily a valuable one.

DECISION EQUATION

net AI value = economic benefit of the work and decisions changed − full cost of execution and control

  • +

    Economic benefit

    Capacity released, cycle shortened, conversion improved, loss avoided, lower cost to serve or revenue enabled.

  • −

    Model and inference

    Model consumption, context, retrieval, evaluation and AI infrastructure.

  • −

    Integration and operation

    APIs, orchestration, observability, data pipelines, maintenance and reliability engineering.

  • −

    Human supervision

    Review, exception handling, quality assurance, investigation and operational oversight.

  • −

    Error and risk

    Expected economic cost of wrong decisions, incidents, fraud, remediation and control failures.

The business case is built around the unit of work or decision that changes — not around tokens, prompts or model accuracy in isolation.

Two of those five lines are people, not infrastructure. An agent handed over at no cost still carries supervision and the expected cost of its errors, and that is the part a CFO will find whether or not the page states it.

APACHE 2.0

What we hand over, and what we do not claim.

Handed over in full, with no conditions

Each agent's method, the schema of its records, the code that computes, its acceptance criteria and the way to measure them.

What we invite

  • Use it inside your organization without asking permission, and adapt it to how you work.
  • Change the criteria: the thresholds are configuration, not code.
  • Build the capability you are missing. If you practise a function that is not on the list, its agent is written by whoever knows the work, not by whoever knows how to program.
  • If you find it wrong, open an issue with the document that proves it. That is worth more than a star.

What we do not claim

  • This produces working drafts, not decisions.
  • Nobody has certified anything here.
  • The agents know what was written down, not what was said outside the documents.

Installation

The repository is public at github.com/josenanez/criterio. It is in testing (0.3): reliability figures are published when they are measured.

Claude Cowork

Customize → Explore plugins → Personal → + → Add marketplace from GitHub → josenanez/criterio

Claude Code

/plugin marketplace add josenanez/criterio /plugin install criterio-portfolio@criterio /plugin install criterio-project@criterio /plugin install criterio-product@criterio

Install only the agent for your role: Vera for the portfolio, Samuel for a project, Alba for a product. Each one has a setup command that looks at your folders, asks a few questions and produces a first result on your own documents. Nobody edits a configuration file by hand.

CONTINUE THE THREAD

Strategy→The thesis these agents are the proof of.Products→Where operating leverage is measured, product by product.Governance→The six conditions this code is built against.