The Workflow Intelligence

Manifesto

We are building AI wrong.

Not the models. Not the frameworks. Not the infrastructure.

The workflows.

Every day, organizations across every industry are deploying AI workflows designed in a rush, executed without review, priced without visibility, and governed without accountability. They are automating at speed while flying blind — and calling it transformation.

This is not transformation. This is acceleration without intelligence.

We need to talk about what's missing.

The Layer Nobody Built

The AI ecosystem is extraordinary. We have foundation models of staggering capability. We have open-source orchestration frameworks that can chain agents, manage memory, and coordinate multi-step reasoning. We have deployment infrastructure that can scale to millions of users overnight.

What we do not have is the layer between intent and execution.

The layer that answers the questions every operator, developer, and founder should be asking before they run a single workflow:

What will this actually cost?

Is this the right model for this task?

Does this workflow reflect the SOP it was built from — or did something get lost in translation?

If this runs a thousand times a day, what does that look like at the end of the month?

Who approved this? Who can change it? Who is accountable when it fails?

Nobody built this layer. So organizations skip it. They go from SOP to execution, from idea to deployment, from intention to invoice — without anything in between.

The gap is not technical. The gap is architectural.

What We Got Wrong About AI Workflows

We inherited a mental model from software development: build it, ship it, monitor it.

It served us well when the cost of a bug was a bad user experience. It does not serve us when the cost of a poorly constructed workflow is tens of thousands of dollars in unnecessary compute, a regulatory violation, or a patient outcome.

We treated AI workflows like code — things to be written and run. We forgot that a workflow is also a policy. A commitment. A representation of institutional knowledge that took years to accumulate and minutes to corrupt.

The standard operating procedure sitting in your SharePoint folder is not just documentation. It is your organization's intelligence about how work gets done, encoded in language. When you hand that to an AI system without compiling it — without translating it into a structured, validated, cost-aware specification — you are not leveraging institutional knowledge. You are gambling with it.

There are millions of SOPs sitting in enterprise SharePoint folders right now that have never been compiled into an AI-ready format. They will never be used. Or worse — they will be used badly.

Introducing SKILL.md

A SKILL.md file is a structured, plain-text specification that encodes a single operational skill — a discrete unit of work that an AI system can understand, cost, and execute.

It is not a prompt. It is not a script. It is not a configuration file for a specific platform.

It is a portable, platform-agnostic standard that captures:

• What the skill does — its purpose, scope, and intended outcome
• How it does it — the step-by-step logic, in structured form
• What it costs — the model, token estimate, and execution mode required
• What governs it — who owns it, when it was approved, what policy it reflects
• Where it runs — which frameworks, models, and deployment targets it is compatible with

A SKILL.md file can be read by a developer, reviewed by a compliance officer, validated by a QA system, and executed by any compatible AI framework — without modification.

It is the missing interface between human knowledge and AI execution.

SKILL.md is an open specification. It does not belong to Cromus. It belongs to the community that adopts it, extends it, and builds with it. Our commitment is to maintain the spec, publish the validator, and never lock it behind a proprietary wall.

You can use SKILL.md today, with or without Cromus, in any AI workflow you build.

Introducing ETHOS.md

SKILL.md answers the question: what does this workflow do, and what will it cost?

That was necessary. It was not sufficient.

There is a second question — one the industry has been avoiding with remarkable consistency — that every organization deploying AI agents must eventually answer:

How is this agent supposed to behave?

What values govern its decisions when the instructions run out?

Who reviewed it? Who approved it? Who is accountable when it acts in a way nobody anticipated?

Is this agent operating within the regulatory and ethical boundaries our industry requires — or did we just assume it would?

These are not engineering questions. They are governance questions. And until now, the answer to all of them has been the same: hope.

ETHOS.md is the behavioral governance specification for AI agents. It is the companion standard to SKILL.md — where SKILL.md encodes what a workflow does, ETHOS.md encodes how an agent is permitted to do it.

It captures:

• Behavioral profile — the agent's role, operational domain, and declared purpose
• Permitted and prohibited actions — explicit boundaries on what the agent may and may not do
• Escalation paths — what happens when the agent encounters ambiguity, conflict, or edge cases beyond its specification
• Compliance coverage — the regulatory frameworks, industry standards, and internal policies the agent is designed to satisfy
• Governance metadata — who owns this agent, when it was reviewed, when it expires, and who has authority to change it

An ETHOS.md file can be read by a developer, reviewed by a compliance officer, validated by a governance tool, and cited in a regulatory audit — without modification.

It is not a prompt. It is not a system message. It is not a policy document that lives in a folder nobody opens. It is a machine-readable, human-auditable behavioral contract — attached to the agent, versioned with the agent, and reviewable by everyone accountable for the agent's outcomes.

Like SKILL.md, ETHOS.md is an open specification. It does not belong to Cromus. It belongs to every organization that has ever deployed an AI agent and quietly wondered whether they had done enough to govern it well.

The answer, until now, was usually no. ETHOS.md is the infrastructure for a better answer.

Introducing MEMORY.md

MEMORY.md is the open specification for portable AI agent memory. MEMORY.md defines what an agent retains across sessions, how it compresses context when limits are reached, and how memory transfers between agents. It works universally — at the model level, the infrastructure level, and the application level — with no proprietary tooling required.

The MEMORY.md specification is open, versioned, and available at:

github.com/cromus-ai/memory-md-spec

Together, SKILL.md, ETHOS.md, MEMORY.md, and GRAPH.md form the complete open agent specification stack. Four files. Everything an AI agent needs to know what to do, how to behave, what to remember, and how capabilities connect — portable, auditable, and platform-agnostic.

Defining a New Category: Workflow Intelligence as a Service

Workflow automation asks: how do we make work happen faster?

Workflow intelligence asks: how do we make work happen smarter — before we spend a dollar making it happen at all?

These are different questions. They require different platforms.

Workflow Intelligence as a Service is a platform category defined by a single governing principle: intelligence precedes execution.

A Workflow Intelligence platform does not run your AI workflows. It makes them ready to run — optimized, costed, governed, and portable across whatever execution environment you choose.

It does four things that no automation platform does:

This is Workflow Intelligence as a Service — a category we are defining here, with this document, for the first time. Cromus is the first platform built expressly to deliver it.

What We Believe

A Call to the Community

If you are a developer building AI workflows: adopt SKILL.md. Write your skills in a format that travels with you, that your teammates can read, that your organization can audit.

If you are a founder building in the AI space: define your workflows before you deploy them. Cost them. Govern them. Treat them as the institutional assets they are.

If you are an operator bringing AI into an enterprise: demand portability. Demand cost visibility. Demand that the intelligence layer be separate from the execution layer so you are never locked into a vendor you cannot leave.

If you are an educator teaching the next generation of AI practitioners: teach workflow design as a discipline. Teach that the quality of an AI workflow is not determined at runtime. It is determined in the design phase — and design requires intelligence before execution.

The AI community built extraordinary tools for execution. It is time to build the intelligence layer that makes execution worthy of the tools.

The Specifications

The SKILL.md specification is open, versioned, and available at:

github.com/cromus-ai/skill-md-spec

The ETHOS.md specification — the behavioral governance standard for AI agents — is open, versioned, and available at:

github.com/cromus-ai/ethos-md-spec

The MEMORY.md specification — the portable AI agent memory standard — is open, versioned, and available at:

github.com/cromus-ai/memory-md-spec

The free SKILL.md, ETHOS.md, MEMORY.md, and GRAPH.md Validators — to score, profile-match, and audit any specification — are live at:

cromus.ai/validator

The Cromus platform — the first Workflow Intelligence as a Service implementation — is live at:

cromus.ai

Beyond the Platform

Cromus is not just infrastructure. It is a founding argument.

The organizations winning with AI are not the ones with the most models — they are the ones who understand what their workflows actually cost, where they fail, and why. That knowledge has been locked inside vendor dashboards, opaque pricing tiers, and proprietary tooling long enough.

Cromus exists to change that. Not just by building better tooling, but by building a community that shares and grows collective knowledge around AI workflows — openly, transparently, and in public.

We believe AI companies should be held to pricing transparency. That teams deserve to know their Total Cost of Workflow Ownership before they're locked in. That the practitioners doing the work should own the intelligence about how that work runs.

SKILL.md, ETHOS.md, MEMORY.md, and GRAPH.md are open specs because the foundation has to be open. The community is the compounding advantage — not any single platform.

We are building both.

Workflow Intelligence as a Service is a category, not a product. The category belongs to the community. The responsibility for defining it well belongs to all of us.

— Alva Cardona

Orlando, Florida — March 2026

We got the sequencing wrong.

In the original manifesto, we argued that AI workflows need a pre-execution intelligence layer — a place to cost, govern, and validate before a single token is spent. That argument holds.

What we did not say clearly enough — and what the product failed to reflect until now — is who deserves that layer first.

We assumed governance was a team problem. It is not. It is a practice problem. And it starts with the first person who writes a SKILL.md file.

The Assumption We Buried in the Pricing Page

Look at the original Cromus plan structure and you will find a quiet assumption embedded in the tiers: the Specialist — the role that reviews every SKILL.md change in CI, that enforces the 24 rules before a workflow merges, that stops the bad workflow before it runs — was locked at the Ops plan. A team plan. A $99 plan.

That meant a solo developer, an indie hacker, a founder shipping their first agentic product alone — the people most likely to have no one else reviewing their work — were the people we decided didn't need the Specialist.

They got cost simulation. They got model comparison. They got scoring.

They did not get governance.

We told ourselves this was about team workflows needing CI enforcement more than solo workflows did. That was rationalization. The truth is simpler: we drew the tier line in the wrong place.

The Department Nobody Staffed

Every organization building with AI needs the same three things, regardless of size:

These are not features. They are roles. And they map — not coincidentally — to how mature organizations have always structured operations functions: the Specialist who does the review, the Manager who watches the portfolio, the Chief Workflow Officer who owns the governance posture.

The AI department has always been the right mental model. What changed is when it starts.

What Shifted — and Why

In August 2026, we moved the Specialist to the Builder tier. Every paid Cromus account — starting at $49/month — now activates a Specialist in CI. Every pull request that touches a SKILL.md is reviewed automatically. Every merge that would introduce a governance violation is blocked before it lands.

This is not a feature unlock. It is a philosophical correction.

The Specialist belongs at the entry tier because the entry-tier user is the one writing the workflows that will scale. The governance habits built in the first hundred SKILL.md files determine the governance posture of the first hundred-person team. We cannot fix bad foundations with better team tooling. The foundation has to be right from the first commit.

The Manager joins at Ops — for teams who need portfolio oversight and shared workspaces.

The CWO joins at Agency — for practitioners who deliver governed AI workflows to clients.

But the Specialist is there from the beginning, for everyone, because the review that matters most is the one before anything ships.

The department scales with you. It does not ask you to earn governance.

A Note on Founding Members

Founding Members joined Cromus early, at $29/month, locked forever. Some joined expecting the Ops-level feature set that was implied when they signed up — team seats, shared libraries, version history. That set changed when the product shifted.

We owe them honesty about what that means: Founding Members have Builder-level features at $29. They have the Specialist, the CI Gate, the full simulation and cost intelligence layer. They do not have the Manager, the team workspace, or the governance report layer. Those belong to Ops and above.

What Founding Members have that no other plan offers: a permanent price lock, the earliest badge in the community, and the right to shape the product roadmap in ways that later cohorts won't. The lock is unconditional. The price never changes regardless of where the Builder tier moves.

We made a product decision. Founding Members deserved to hear it directly, not discover it on a pricing page. This is us saying it directly.

What We Now Believe

The New Call

If you are building your first AI-powered product alone: you now have a Specialist in your CI pipeline from day one. Use it. Write the SKILL.md file. Let the Specialist review it. Build the governance habit before you have a team that needs it.

If you are leading an AI team: the Manager is waiting. Portfolio visibility, shared libraries, team workspaces — the infrastructure to coordinate the practice you already built alone.

If you are an agency or consultancy delivering AI workflows to clients: the CWO is the difference between a project and a practice. Governance reports, behavioral attestations, client-facing sign-off — the full department, under your brand.

If you are an enterprise: bring us in at the scale your compliance function requires. Custom seats, dedicated support, the full department plus the contract your legal team needs.

The department is there. It staffs automatically as you grow. It never asks you to earn governance — because governance is where the department begins.

The department before the headcount. The review before the merge. The governance before the problem.

— Alva Cardona

Orlando, Florida — August 2026