Everyone I discuss it with agrees with the asymmetric operating model concept, but nobody has actually built it. Not for lack of vision. For lack of execution capacity.
Three earlier pieces set the frame:
- Part 1: asymmetric supply chains obey different physics
- Part 2: the operating model must match
- Part 3: discusses how chaos concepts shape our performance.
This is a closing one. It comes out of live work with my teams - deploying agentic capabilities, watching what happens, revising what I thought I understood. Every quarter produces more questions than answers, but also more clarity about which questions matter. What follows is what I have come to see clearly enough.
Why We Never Fixed This
There is a simple rule that quietly explains fifty years of failed transformation programs in agri-food.
A management system can only handle situations as varied as itself is. Give it a hundred kinds of problem and it does well. Give it a thousand - it recognizes a hundred, and the other nine hundred just happen. To you.
Now look at what we face in our supply chain context.
The environment: 12.5 million smallholder coffee farmers, micro-climates that shift with each rainy season, El Niño years, port strikes, government export tariffs, market algorithms deciding your basis before you finish breakfast. Effectively unbounded.
The management system: one central ERP, weekly planning, a static supplier master, exception reports.
We built control rooms with the range of an assembly line and pointed them at an environment with the range of a weather system. It was never going to work
Every response we tried made it worse two ways. Reduce the environment's variety - standardized specs, rigid contracts, force chaotic origins to behave like automotive suppliers, crush the fragmented farmer base and lose the upside volatility creates.
Or centralize harder - more dashboards, more control towers, more visibility, and still no one able to act on any of it at origin speed.
Visibility without matching response capacity is just a better view of the crash
Where Agents Actually Help
Three things. At least the way I see this now, but it might change:
1. Judgement at origin. One agent per origin, per local coffee flow corridor, per shipping lane. Watching yield anomalies, monitoring quality drift, tracking logistics friction. Interpreting local conditions continuously.
Practically: origin team starts morning to find yesterday's export documents already checked, yesterday's price signals from three cooperatives already summarized, and the three suppliers whose recent lots drifted on screen size already flagged. That is what agents change. As long as the interpretation is actually good, which is not a given.
2. Cheaper optionality - not free. Backup origins, alternate corridors, postponed commitments used to decay quietly. Every dormant supplier needs qualification updates, relationship maintenance, spec re-checks. Nobody had time. So the fifty-supplier map that looked good in a slide became a five-supplier map by year-end.
Agents cut the cost of keeping the other forty-five warm. Re-validating, re-scoring, re-pricing them continuously. But cheaper is not free. There are running costs. There is maintenance. And cheaper does not mean reliable. Watching is easy. Interpreting is where these programs actually fail.
3. Same clock as the market. A frost in Minas Gerais hits paper in hours. If your physical response cycle is three weeks, you receive that cascade. You do not participate in it. Agents let the managed response run on the same clock as the emergent behavior - for the decisions where speed matters. For decisions where it does not, machine speed is expensive theatre.
What Agents Do Not Fix
In my opinion reaching for agents everywhere is the ERP mistake, second edition. Parts of the model do not need them:
- Federated authority is an organizational choice. Distributing decision rights to origin-country teams do not require AI. Most companies simply refuse to do it. The blocker sometimes is politics, not technology.
- Parametric risk - weather derivatives, index insurance, El Niño coverage - is contracts and data. It needs willing counterparties and better hydrological indices. Not intelligence.
- Postponement, dual origin, safety stocks are network design choices. Rules-based systems and better contracts execute most of it. Agents help at the margin.
Trust primitives - origin traceability, quality attestation, EUDR-compliant provenance - need shared standards and interoperable systems across the industry. Not smarter software inside one firm.
The failure of the last two decades in agri-food was not that we lacked agents. The operating model was pointed the wrong way. Agents accelerate a correctly-pointed model. They accelerate the wrong direction just as fast.
And they carry costs of their own. Oversight, testing, escalation handling, the humans who used to make those calls slowly losing the muscle. Nothing is free.
Precondition: if your operating model cannot delegate a single meaningful decision, no agent will change that. The organizational blocker precedes the technology decision. Fix the first, or the second is theatre.
What Has To Change First
Put agents inside a centralized command-and-control model and you have built a faster ruler. Every decision still queues behind the same head-office bottleneck. You get the cost of agents with the responsiveness of the old model.
This triggers three design decisions that are also not easy depending on legacy culture and setup of your operating model:
Decide who decides - by cost of being wrong, not by seniority. Delegate what is frequent, local, reversible, bounded. Rerouting containers. Re-scoring a supplier after a bad lot. Adjusting sample submissions. Reserve what is rare, irreversible, strategic. Country entry. Major forward positions. Ending a fifteen-year cooperative relationship.
The question is never "can the agent decide?" It is "what does it cost when it decides wrongly?"
Autonomy at origin. Policy at head office. Coordination in between. Most programs treat the middle layer as an afterthought - how do the different agents actually talk to each other, resolve conflicts, avoid working against one another? Then they pay for it in the first bad quarter, when the risk agent is buying coverage the sourcing agent has already made unnecessary.
Human judgement as a designed dial. Per decision class. Explicit. Revisited as trust builds. Humans in the loop for judgement, not routine. Most organizations never set the dial deliberately - it gets set by default fear, or by default enthusiasm, which is worse.
The Agent Monoculture Problem
The same rule that indicts centralized ERPs also indicts the agent fleet.
A thousand agents all built on the same underlying AI, trained on the same data, watching the same signals, running the same playbooks - that is not a thousand units of judgement. It is one, repeated a thousand times.
Counting agents is the wrong metric. Genuinely independent reasoning is the right one.
Agent design is a portfolio decision, not a purchase decision. Buy from one vendor, run one AI across your whole operation, and discover the problem the first time the shock hits everyone at once.
When Everyone's Agents React The Same Way
This is the failure mode agri-food has never faced. And here I speculate a bit trying to imagine the situation when:
Every trading house, processor, and logistics operator running similar agents, watching similar signals, drawing similar conclusions. A local shock - Suez closure, Brazil frost, Vietnam typhoon - triggers a synchronized, machine-speed response across the whole industry.
Financial markets ran this experiment in 2010. The flash crash was not caused by bad algorithms. It was caused by identical ones. Automated systems pausing on the same thresholds at the same moment.
Financial markets had circuit breakers, market makers, and a regulator. Coffee has none of that. Cocoa has none. Cotton has none. When the first machine-speed bullwhip rips through a soft commodity, there is nobody to call.
The firm-level circuit breakers this piece calls for - hard limits on autonomous action, escalation on industry-wide anomalies, deliberate diversity across the agent portfolio - are the substitute for market-level infrastructure that does not exist.
Design for it now, or explain it later...