Nine agents. Forty-two grams.

Formulation is an argument nobody wins

Designing a supplement means satisfying constraints that actively fight each other. The protein claim on the front of the pack sets a floor. The scoop volume sets a ceiling. Flavour needs mass to work at all. The regulator has views on every ingredient. The accountant has views on all of it. Stability decides whether any of it survives a Pakistani summer in a warehouse.

No single person holds all of that. In practice it gets resolved by whoever is most senior in the room, in a spreadsheet nobody else can audit, and the reasoning disappears the moment the decision is made.

The interesting part is not that the constraints are hard. It is that they are competing, and the trade-off is usually invisible.

Make the budget physical, then make them bid for it

The whole system rests on one decision: the scoop is 42.00 grams and that number is not negotiable. It is a physical object. It does not care about anyone's priorities.

So the agents do not discuss. They bid. Each specialist submits a claim on grams of the budget, and what remains is calculated and published as residual headroom for everyone to see.

Because the budget is zero sum, no agent can win without another losing, and every concession is recorded rather than absorbed. The flavour agent bidding its mass down below the conventional load is not a preference, it is a documented trade made to fund the protein claim.

That is the difference between a system that produces a decision and a system that produces an auditable decision.

Nine specialists, deliberately in conflict

Each has a single remit and no obligation to be agreeable about it.

Kernel

Orchestrator, custodian of the single source of truth, and arbitrator when two agents deadlock.

Nexus

The protein matrix and clinical dosing. Holds the floor that the front-of-pack claim depends on.

Chroma

Flavour, sweetness and masking, working against a mass budget rather than a recipe.

Forge

Rheology, blend sequencing, and which machinery the formulation actually requires. Equipment follows the formulation, not the other way round.

Stasis

Stability under ICH Zone IVb, moisture ingress, and how much overage the claim needs to survive shelf life.

Vector

Landed cost, the local tax stack, and a hard margin floor beneath which a formulation is rejected on economics alone.

Protocol

Regulatory compliance, banned-substance screening, and certification audit.

Apex

Innovation and arbitrage of whatever headroom is left after everyone else has taken their share.

Zero

Red team. Its only job is to attack the result the other eight agreed on.

Four phases, and a kill switch

State lives in one locked JSON object rather than in conversation. It carries the phase, a conflict register, loop counters, and deadlock flags that any agent can raise to stop the run.

The kill switch is the part most agent systems lack. If the same disagreement circulates three times, the run halts and escalates to a human instead of negotiating forever. Agents that cannot fail visibly will burn budget until someone notices, which is the standard way these systems go wrong.

The stability agent can also raise a hard deadlock, on the grounds that a formulation which will not survive its own shelf life is not a formulation worth costing.

What a run actually produces

Not a chat transcript. Fourteen structured specification files, each one machine readable and human auditable, covering formulation, flavour, manufacturing, stability, cost, compliance and risk.

• Specialist agents in the cluster9
• Scoop budget, fixed42.00 g
• Residual headroom after all bids0.07 g
• Specification artefacts per run14
• Blend sequence, steps and cycle time5 · 65 min
• Formulation ratios, cost and marginwithheld

The protein claim landed on the target as a literal figure rather than a rounded one, which matters because rounding a claim upward is a regulatory problem rather than a marketing one.

An agent whose only job is to attack the answer

Zero takes the agreed specification and runs failure mode and effects analysis against it, scoring severity, probability and detectability, and ranking by risk priority number against fixed thresholds.

This is the part that transfers furthest outside nutrition. Most automated decisions are never adversarially tested before they ship.

What this is, and is not

Built for a founder-led sports nutrition company in Pakistan, working directly with its co-founder across formulation strategy, landed cost and regulatory compliance.

The output is a specification, not a lab-validated product. The system reasons about formulation, stability and compliance, and its conclusions are internally consistent and traceable, but computed chemistry is not tested chemistry. Anything intended for sale would need laboratory assay and real stability data behind it. In food and supplements that distinction is not a technicality.

Formulation ratios, landed cost and margin are withheld here. The transferable part was never the recipe.

Where this shape turns up again

Any answer that has to satisfy constraints which disagree.

A quotation bounded by a code, a margin floor and a customer's budget. A schedule bounded by skills, travel and statutory deadlines. A purchase bounded by lead time, cash and shelf life. In each case the constraints are known, they conflict, and the reasoning normally lives in one experienced person's head where it can be neither checked nor repeated.

The pattern is the same every time: make the budget explicit, make the specialists bid against it, record every concession, and let something adversarial attack the result before it ships.