In 2004 I specified an architecture that required agents capable of acting on real-world operating attributes. For twenty-two years the only things that could do it were people — and a set of self-learning algorithms with nowhere to go.
The argument, in four lines.
1. In 1998 I specified an agent-based architecture. Only humans could occupy the boxes; the algorithms sat in the hub under human monitoring because they had nowhere else to stand.
2. In 2026 that intelligence can finally be distributed into the boxes. Same architecture, constraint removed.
3. But most of what is now drawn as a Metaprise is still one agent’s internals — Rio, and the cloud components poster.
4. Distributed or not, the specification remains unmet until an agent can surface undeclared related attributes. And accountability stays in the hub.
The rest of this post is the evidence for each.
A post came through my feed this week with a diagram titled Cloud Components Required for Modern AI Systems.
A cloud at the center, boxes arranged around it, connecting lines running between them, the logos of the three major providers distributed across the page.
I looked at it for a while and could not make it mean anything. Then I recognized it.
The brochure
In 1999, my company published a brochure built on work from the 1998 engagement. On it was a two-panel diagram.
The original brochure graphic, 1999, reproduced as published.
The left panel is headed Enterprise Applications (Sequential): Supplier, Manufacturer, Reseller, in a straight line, each handing off to the next.
The right panel is headed Metaprise Applications (Synchronous): participant blocks arranged around a central HUB, each connected to it and acting through it.
It is the right-hand panel I recognized. Same hub. Same boxes around it. Same connecting lines.
My first reaction was that someone had taken an operating ecosystem, removed the people, and stacked technology in the empty space.
That was wrong, and what is actually going on took me most of a day to see properly.
What was specified, and when
The architecture was agent-based, and I put the definition in writing in the autumn of 2004, in a paper titled Acres of Diamonds — quoted again in a June 2007 post, four weeks after this blog began:
It is my position that a true centralization of procurement objectives requires a decentralized architecture that is based on the real-world operating attributes of all transactional stakeholders starting at the local or regional level… This is the cornerstone of agent-based modeling.
By 2008 I had stated the contrast plainly. Equation-based models quantify and therefore confine multiple strands — the operating attributes of diverse stakeholders — into a single definable static process. Agent-based models do the opposite: they begin from those attributes as they actually are, then link seemingly disparate ones to produce a real-world outcome in real time.
That is a technical specification, not a philosophy. It says what an agent has to be able to do: hold real-world operating attributes, act on them, and have its actions linked to other agents’ attributes across strands that nobody has declared to be related.
What was in the middle
Here is the part I had been walking past.
The hub was not a router. It was the orchestration position — and in 1999 it held the advanced self-learning algorithms as well, because they had nowhere else to go.
That was the technology: SR&ED-funded R&D, built to detect related attributes across seemingly disparate strands and act on them in real time. The intelligence sits in the center of that diagram, and it sits there out of necessity rather than design.
⭐ But the center was never only technology. It is where the human orchestrator sits — overseeing what amounts to a train yard, with the agents running on their own tracks, carrying their own loads, on their own schedules. The orchestrator does not drive the trains. The orchestrator sees the whole yard, and is answerable for what happens in it.
And this is where the governance mechanism resides. Not as a component in one of the boxes. In the center, with the orchestrators — because governance is the authority to set what the agents may do, to stop a recommendation, to resolve conflicting outcomes, and to accept responsibility when the agents perform correctly and the result is still wrong.
No agent can hold that, however capable. An agent can be governed. It cannot govern.
And the boxes held people — supplier, manufacturer, reseller, buyer, technician, customs broker — because nothing else could hold a role. A machine could not carry its own context, act on its own attributes, or be given a position in the actual external operating environment and left to perform it.
I used the word for it. In an October 2023 post describing how the 1998 engagement was actually run:
One of the first things I did was understand what role the other stakeholders or “agents” played in procurement’s success.
And on why the obvious remedy would not work:
The likelihood that we could change technician or supplier behavior, e.g., “agents” for parts that were required the next day, was virtually impossible.
Note the quotation marks. I was borrowing a term and applying it to people, because people were the only available candidates. Note also the word all in the 2004 definition. The specification was broader than anything that could then meet it.
⭐ So the algorithms sat in the middle, doing agent work, unable to be agents. They could detect relationships across strands, but they could not be distributed. They could not hold a position in the actual external operating environment. They stayed in the hub because 1998 gave them nowhere else to stand.
Now they can be set free
That is what the 2026 diagram actually shows, and why I recognized it.
The intelligence is no longer confined to the center. A model, an agent, a retrieval process can now hold context, act on attributes, and occupy a position in the actual external operating environment. The thing that was in the hub can now be distributed into the boxes.
That is not a different architecture. It is the same architecture, with the constraint removed.
⭐ And notice what is left in the middle when the algorithms move out. The hub does not empty. It returns to being what it always was underneath: the orchestration position, occupied by a human, overseeing a yard that now contains both classes of agent.
The algorithms were tenants in that space. The orchestrator was the resident.
And I am not the only one who has noticed the lineage. In August 2025, writing about ConvergentIS and their Rio procurement agent, I reproduced their own graphic — headed The Evolution of the Agent-Based AI Operating System:
The Metaprise brochure diagram on the left, dated 1999. A hub with distributed agents in the middle, dated 2024. The Rio Procurement Agent on the right.
That lineage was not drawn by me. It was drawn by a solution provider with twenty-three years of SAP delivery behind them, and it places the brochure at the origin of the line. I take that as a generous and well-informed reading.
But look carefully at the third panel
Rio is arranged the way my synchronous panel is arranged — a center with blocks around it. Conversational channel, document channel, messaging channel, enterprise SaaS channel.
Those are not other agents. They are Rio’s own interfaces.
Which means the entire Rio model resides in a single agent block. It is one participant. In the architecture on the left of that same graphic, Rio occupies one box — alongside the supplier, the buyer, the technician, the customs broker.
Now look again at the cloud components diagram.
Models, retrieval, vector search, pipelines, orchestration, APIs, monitoring, guardrails. That is also one agent’s internals, drawn at full page size using the visual grammar of a multi-agent environment.
⭐ Neither diagram is a Metaprise. Both are pictures of a single agent, in the shape of an operating environment.
Which finally explains the thing that bothered me on first reading. I kept looking for the operating participants — the customers, the suppliers, the decision-makers, the competing objectives — and could not find them.
There was never room for them. The whole page is one box.
Twenty-two years from that 2004 definition, and the design did not have to change to accommodate it. It was already agent-based, already synchronous, and already required agents capable of acting on real-world operating attributes. What arrived is a second class of candidate that meets the specification.
And a single agent still has to meet the specification
Being one box is not a criticism. Every agent in that architecture is one box. The supplier was one box. The question is what the agent inside it can do.
Which is where the newer diagram tells on itself.
Look at what it contains: models, agents, retrieval, vector search, pipelines, orchestration, APIs, security, monitoring, cost control — and guardrails.
Guardrails, not firm fences. That distinction is the entire thing.
A guardrail constrains an agent inside a frame someone else drew. It governs what the agent may do, what it may touch, what it must escalate. It is necessary and it is not sufficient, because nothing about a guardrail permits an agent to look at a strand nobody put in scope.
A firm fence backed by advanced self-learning algorithms operates differently. It is not a boundary on permitted action; it is a requirement placed on the agent itself — that it detect related attributes across strands that were never declared to be related, and surface them.
In January 2025 I put the question publicly to a group of people working on exactly this problem:
Tell me how Agentic AI would have known to ask, “What time of day do orders come in?”
The answers described an agent recognizing order-timing patterns in the data — in a dataset someone had already decided to collect, containing a field someone had already decided to record.
That is not the question. So I asked it more precisely:
How does Agentic AI address the seemingly disparate strands of data in which unidentified “related attributes” collectively impact the desired outcome?
Nobody has answered it.
In 1998 the four o’clock question was askable because the architecture was built to surface relationships nobody had declared. The technicians’ call-response measure, the buyers’ queue, the suppliers’ cross-border inexperience, the courier fragmentation, and finance’s two separate line items were five strands with an unidentified related attribute running through them. Every measure was accurate. Nothing in the system connected them.
An agent with guardrails cannot find that, because the guardrail assumes the frame is correct and governs behavior inside it.
⭐ So the boxes in the 2026 diagram are filled with agents that cannot yet do the thing the architecture was specified to require in 1998. The intelligence has been distributed. The specification has not been met.
The asymmetry that does not dissolve
One more thing, and it holds regardless of how good the algorithms get.
As agents, human and machine can be equivalent — and in accountability they are not. Those two clauses belong together, and I would ask that they not be separated.
Each can hold attributes, act on them, and shape an outcome no single agent controls. That is the equivalence, and it is real. A human agent can be asked why, can refuse, can escalate, can say this will not work and carry the consequence of saying it. A machine agent executes inside whatever it is given, and nothing in the exchange can tell you the frame was wrong.
Which is why AI success within an enterprise is not putting humans in the loop. It is putting them at the helm of a system running on either a firm fence backed by advanced self-learning algorithms, or guardrails requiring more direct human intervention — with a transparent audit trail behind both.
And “at the helm” is not a figure of speech here. It is a position in the architecture. The hub. The place the orchestrator has occupied since 1999, which does not become vacant because the algorithms finally have somewhere else to stand.
Which brings the 2004 statement back, and this time the second sentence is the one that matters. Here it is in full, from Acres of Diamonds:
It is my position that a true centralization of procurement objectives requires a decentralized architecture that is based on the real-world operating attributes of all transactional stakeholders starting at the local or regional level. In other words, your organization gains control of its spend environment by relinquishing centralized functional control in favor of operational efficiencies on the front lines. This is the cornerstone of agent-based modeling.
Gains control by relinquishing control. That is not a paradox, and it is not a slogan. It is a description of what the center is for.
The orchestrator does not control the agents. The agents act on their own operating attributes, at the edge, where the work actually happens. What the center holds is the objective, the authority, and the accountability — which is exactly why distributing the algorithms into the boxes does not diminish the hub. It is what the hub was built to make possible.
An architecture with two classes of agent needs one class accountable. Otherwise every participant in the diagram executes correctly and nobody owns the result — which is precisely the failure I traced in 1998, five departments deep.
That engagement moved next-day delivery from 51% to 97.3% in three months. Cost of goods fell 23% year over year for seven consecutive years. The buyer FTE count went from 23 to 3 inside eighteen months. The technology went live in production in August 2003.
The groundwork was done before the technology was introduced. It always is.
What a components map still cannot do
None of this rescues that diagram as an architecture, and it is worth being precise about why.
It has no left panel. My brochure graphic makes a comparative claim — here is how it works now, here is how it works instead. That claim can be argued with. A components map has no before, so there is nothing to disagree with.
And notice where it puts governance. A box, alongside monitoring and API management, with vendor services listed underneath it. But no service in that box establishes who holds decision authority, who may stop the system, how conflicting outcomes are resolved, or what happens when the technology performs exactly as designed and the organizational outcome still fails.
Governance is not a component sitting beside the others. It sits in the center, with the orchestrators, and it is the condition under which every box is permitted to act.
And here is the test. In a traceback map, remove a box and the explanation breaks — the chain no longer accounts for the outcome. In a components map, remove “containers” or “data warehouse” and nothing changes, because no logic connects the boxes beyond these things may all be involved.
Removing a box shortens the inventory rather than breaking the argument.
The question
If an AI implementation fails in your organization — the agent approves something it should not, the recommendation is confidently wrong, the pilot never reaches production — a components map can tell you which technologies were present.
It cannot tell you which relationship produced the failure, because it does not contain relationships. It contains an inventory.
So three questions, and the last one has been open since January 2025.
Which box do you remove to break the explanation?
When your diagram holds both classes of agent — which one is accountable for the outcome?
And: tell me how Agentic AI would have known to ask what time of day the orders come in.
-30-
⚠ A note on the numbers. This archive spans nineteen years and was written contemporaneously, which means figures occasionally vary slightly between entries — the share of US-based suppliers appears as 85% in one place and roughly 80% in another, because each was written from what was in front of me at the time and the two were never reconciled. I am leaving them as they stand. An archive in which every number matches across two decades is not a contemporaneous record. It is polished recall.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Jon W. Hansen, FCIPS — Procurement Insights | Hansen Models™
Related
The Algorithms Were Always in the Middle. They Were Waiting to Be Set Free.
Posted on September 18, 2026
0
In 2004 I specified an architecture that required agents capable of acting on real-world operating attributes. For twenty-two years the only things that could do it were people — and a set of self-learning algorithms with nowhere to go.
The argument, in four lines.
1. In 1998 I specified an agent-based architecture. Only humans could occupy the boxes; the algorithms sat in the hub under human monitoring because they had nowhere else to stand.
2. In 2026 that intelligence can finally be distributed into the boxes. Same architecture, constraint removed.
3. But most of what is now drawn as a Metaprise is still one agent’s internals — Rio, and the cloud components poster.
4. Distributed or not, the specification remains unmet until an agent can surface undeclared related attributes. And accountability stays in the hub.
The rest of this post is the evidence for each.
A post came through my feed this week with a diagram titled Cloud Components Required for Modern AI Systems.
A cloud at the center, boxes arranged around it, connecting lines running between them, the logos of the three major providers distributed across the page.
I looked at it for a while and could not make it mean anything. Then I recognized it.
The brochure
In 1999, my company published a brochure built on work from the 1998 engagement. On it was a two-panel diagram.
The original brochure graphic, 1999, reproduced as published.
The left panel is headed Enterprise Applications (Sequential): Supplier, Manufacturer, Reseller, in a straight line, each handing off to the next.
The right panel is headed Metaprise Applications (Synchronous): participant blocks arranged around a central HUB, each connected to it and acting through it.
It is the right-hand panel I recognized. Same hub. Same boxes around it. Same connecting lines.
My first reaction was that someone had taken an operating ecosystem, removed the people, and stacked technology in the empty space.
That was wrong, and what is actually going on took me most of a day to see properly.
What was specified, and when
The architecture was agent-based, and I put the definition in writing in the autumn of 2004, in a paper titled Acres of Diamonds — quoted again in a June 2007 post, four weeks after this blog began:
By 2008 I had stated the contrast plainly. Equation-based models quantify and therefore confine multiple strands — the operating attributes of diverse stakeholders — into a single definable static process. Agent-based models do the opposite: they begin from those attributes as they actually are, then link seemingly disparate ones to produce a real-world outcome in real time.
That is a technical specification, not a philosophy. It says what an agent has to be able to do: hold real-world operating attributes, act on them, and have its actions linked to other agents’ attributes across strands that nobody has declared to be related.
What was in the middle
Here is the part I had been walking past.
The hub was not a router. It was the orchestration position — and in 1999 it held the advanced self-learning algorithms as well, because they had nowhere else to go.
That was the technology: SR&ED-funded R&D, built to detect related attributes across seemingly disparate strands and act on them in real time. The intelligence sits in the center of that diagram, and it sits there out of necessity rather than design.
⭐ But the center was never only technology. It is where the human orchestrator sits — overseeing what amounts to a train yard, with the agents running on their own tracks, carrying their own loads, on their own schedules. The orchestrator does not drive the trains. The orchestrator sees the whole yard, and is answerable for what happens in it.
And this is where the governance mechanism resides. Not as a component in one of the boxes. In the center, with the orchestrators — because governance is the authority to set what the agents may do, to stop a recommendation, to resolve conflicting outcomes, and to accept responsibility when the agents perform correctly and the result is still wrong.
No agent can hold that, however capable. An agent can be governed. It cannot govern.
And the boxes held people — supplier, manufacturer, reseller, buyer, technician, customs broker — because nothing else could hold a role. A machine could not carry its own context, act on its own attributes, or be given a position in the actual external operating environment and left to perform it.
I used the word for it. In an October 2023 post describing how the 1998 engagement was actually run:
And on why the obvious remedy would not work:
Note the quotation marks. I was borrowing a term and applying it to people, because people were the only available candidates. Note also the word all in the 2004 definition. The specification was broader than anything that could then meet it.
⭐ So the algorithms sat in the middle, doing agent work, unable to be agents. They could detect relationships across strands, but they could not be distributed. They could not hold a position in the actual external operating environment. They stayed in the hub because 1998 gave them nowhere else to stand.
Now they can be set free
That is what the 2026 diagram actually shows, and why I recognized it.
The intelligence is no longer confined to the center. A model, an agent, a retrieval process can now hold context, act on attributes, and occupy a position in the actual external operating environment. The thing that was in the hub can now be distributed into the boxes.
That is not a different architecture. It is the same architecture, with the constraint removed.
⭐ And notice what is left in the middle when the algorithms move out. The hub does not empty. It returns to being what it always was underneath: the orchestration position, occupied by a human, overseeing a yard that now contains both classes of agent.
The algorithms were tenants in that space. The orchestrator was the resident.
And I am not the only one who has noticed the lineage. In August 2025, writing about ConvergentIS and their Rio procurement agent, I reproduced their own graphic — headed The Evolution of the Agent-Based AI Operating System:
The Metaprise brochure diagram on the left, dated 1999. A hub with distributed agents in the middle, dated 2024. The Rio Procurement Agent on the right.
That lineage was not drawn by me. It was drawn by a solution provider with twenty-three years of SAP delivery behind them, and it places the brochure at the origin of the line. I take that as a generous and well-informed reading.
But look carefully at the third panel
Rio is arranged the way my synchronous panel is arranged — a center with blocks around it. Conversational channel, document channel, messaging channel, enterprise SaaS channel.
Those are not other agents. They are Rio’s own interfaces.
Which means the entire Rio model resides in a single agent block. It is one participant. In the architecture on the left of that same graphic, Rio occupies one box — alongside the supplier, the buyer, the technician, the customs broker.
Now look again at the cloud components diagram.
Models, retrieval, vector search, pipelines, orchestration, APIs, monitoring, guardrails. That is also one agent’s internals, drawn at full page size using the visual grammar of a multi-agent environment.
⭐ Neither diagram is a Metaprise. Both are pictures of a single agent, in the shape of an operating environment.
Which finally explains the thing that bothered me on first reading. I kept looking for the operating participants — the customers, the suppliers, the decision-makers, the competing objectives — and could not find them.
There was never room for them. The whole page is one box.
Twenty-two years from that 2004 definition, and the design did not have to change to accommodate it. It was already agent-based, already synchronous, and already required agents capable of acting on real-world operating attributes. What arrived is a second class of candidate that meets the specification.
And a single agent still has to meet the specification
Being one box is not a criticism. Every agent in that architecture is one box. The supplier was one box. The question is what the agent inside it can do.
Which is where the newer diagram tells on itself.
Look at what it contains: models, agents, retrieval, vector search, pipelines, orchestration, APIs, security, monitoring, cost control — and guardrails.
Guardrails, not firm fences. That distinction is the entire thing.
A guardrail constrains an agent inside a frame someone else drew. It governs what the agent may do, what it may touch, what it must escalate. It is necessary and it is not sufficient, because nothing about a guardrail permits an agent to look at a strand nobody put in scope.
A firm fence backed by advanced self-learning algorithms operates differently. It is not a boundary on permitted action; it is a requirement placed on the agent itself — that it detect related attributes across strands that were never declared to be related, and surface them.
In January 2025 I put the question publicly to a group of people working on exactly this problem:
The answers described an agent recognizing order-timing patterns in the data — in a dataset someone had already decided to collect, containing a field someone had already decided to record.
That is not the question. So I asked it more precisely:
Nobody has answered it.
In 1998 the four o’clock question was askable because the architecture was built to surface relationships nobody had declared. The technicians’ call-response measure, the buyers’ queue, the suppliers’ cross-border inexperience, the courier fragmentation, and finance’s two separate line items were five strands with an unidentified related attribute running through them. Every measure was accurate. Nothing in the system connected them.
An agent with guardrails cannot find that, because the guardrail assumes the frame is correct and governs behavior inside it.
⭐ So the boxes in the 2026 diagram are filled with agents that cannot yet do the thing the architecture was specified to require in 1998. The intelligence has been distributed. The specification has not been met.
The asymmetry that does not dissolve
One more thing, and it holds regardless of how good the algorithms get.
As agents, human and machine can be equivalent — and in accountability they are not. Those two clauses belong together, and I would ask that they not be separated.
Each can hold attributes, act on them, and shape an outcome no single agent controls. That is the equivalence, and it is real. A human agent can be asked why, can refuse, can escalate, can say this will not work and carry the consequence of saying it. A machine agent executes inside whatever it is given, and nothing in the exchange can tell you the frame was wrong.
Which is why AI success within an enterprise is not putting humans in the loop. It is putting them at the helm of a system running on either a firm fence backed by advanced self-learning algorithms, or guardrails requiring more direct human intervention — with a transparent audit trail behind both.
And “at the helm” is not a figure of speech here. It is a position in the architecture. The hub. The place the orchestrator has occupied since 1999, which does not become vacant because the algorithms finally have somewhere else to stand.
Which brings the 2004 statement back, and this time the second sentence is the one that matters. Here it is in full, from Acres of Diamonds:
Gains control by relinquishing control. That is not a paradox, and it is not a slogan. It is a description of what the center is for.
The orchestrator does not control the agents. The agents act on their own operating attributes, at the edge, where the work actually happens. What the center holds is the objective, the authority, and the accountability — which is exactly why distributing the algorithms into the boxes does not diminish the hub. It is what the hub was built to make possible.
An architecture with two classes of agent needs one class accountable. Otherwise every participant in the diagram executes correctly and nobody owns the result — which is precisely the failure I traced in 1998, five departments deep.
That engagement moved next-day delivery from 51% to 97.3% in three months. Cost of goods fell 23% year over year for seven consecutive years. The buyer FTE count went from 23 to 3 inside eighteen months. The technology went live in production in August 2003.
The groundwork was done before the technology was introduced. It always is.
What a components map still cannot do
None of this rescues that diagram as an architecture, and it is worth being precise about why.
It has no left panel. My brochure graphic makes a comparative claim — here is how it works now, here is how it works instead. That claim can be argued with. A components map has no before, so there is nothing to disagree with.
And notice where it puts governance. A box, alongside monitoring and API management, with vendor services listed underneath it. But no service in that box establishes who holds decision authority, who may stop the system, how conflicting outcomes are resolved, or what happens when the technology performs exactly as designed and the organizational outcome still fails.
Governance is not a component sitting beside the others. It sits in the center, with the orchestrators, and it is the condition under which every box is permitted to act.
And here is the test. In a traceback map, remove a box and the explanation breaks — the chain no longer accounts for the outcome. In a components map, remove “containers” or “data warehouse” and nothing changes, because no logic connects the boxes beyond these things may all be involved.
Removing a box shortens the inventory rather than breaking the argument.
The question
If an AI implementation fails in your organization — the agent approves something it should not, the recommendation is confidently wrong, the pilot never reaches production — a components map can tell you which technologies were present.
It cannot tell you which relationship produced the failure, because it does not contain relationships. It contains an inventory.
So three questions, and the last one has been open since January 2025.
Which box do you remove to break the explanation?
When your diagram holds both classes of agent — which one is accountable for the outcome?
And: tell me how Agentic AI would have known to ask what time of day the orders come in.
-30-
⚠ A note on the numbers. This archive spans nineteen years and was written contemporaneously, which means figures occasionally vary slightly between entries — the share of US-based suppliers appears as 85% in one place and roughly 80% in another, because each was written from what was in front of me at the time and the two were never reconciled. I am leaving them as they stand. An archive in which every number matches across two decades is not a contemporaneous record. It is polished recall.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Jon W. Hansen, FCIPS — Procurement Insights | Hansen Models™
Share this:
Like this:
Related