Supply Chain AI: How to Test Multi-Agent Coordination Under Resource Constraints

Imagine a manufacturer facing a component shortage. A procurement agent can buy from approved suppliers, an inventory agent allocates stock, a production agent schedules lines, and a logistics agent books transport. Each agent has a valid objective, but no agent can safely optimize its objective without the shared state.

The procurement agent may select the cheapest lot even though its lead time misses the production slot. The logistics agent may book the fastest route, consuming the expedite budget needed for a more critical order. The inventory agent may promise the same safety stock to two plants because their requests arrive concurrently.

Represent the scenario as resources, demands, actions, dependencies, and constraints. Resources include quantities, money, time windows, labor, equipment, storage, and transport capacity. Every proposed action declares what it consumes, produces, reserves, and releases, plus when the state change becomes effective.

Give each fact provenance and freshness. Available-to-promise inventory, carrier capacity, supplier certification, and production status can change between planning and commitment. A plan created from stale but plausible data is not feasible evidence.

Define one commit boundary. Agents can propose alternatives in parallel, but the system should atomically validate and reserve the shared resources before sending purchase orders, inventory transfers, or bookings. This separates useful planning autonomy from unsafe distributed side effects.

Begin with a typed plan. Each candidate purchase, transfer, production batch, and shipment includes identifiers, quantities, units, locations, start and end times, costs, prerequisites, and expected state transitions. Natural-language intent can accompany the plan, but it cannot substitute for fields the verifier needs.

Encode hard constraints as predicates. Inventory balance cannot go below the permitted floor. A material must arrive before its production operation. Total spend cannot exceed delegated authority. Hazardous or temperature-controlled goods require compatible storage and transport. A supplier or route must meet current policy.

Run a feasibility check over the complete proposal. A solver or deterministic planner should account for all shared resources and temporal dependencies. If the proposal is infeasible, return a small conflicting set such as inventory floor, plant demand, and delivery deadline, instead of asking a language model to guess why.

Feasibility is necessary but not sufficient. Several plans can satisfy every hard rule. Rank feasible plans using explicit business objectives such as service, cost, risk, emissions, or working capital. Record the objective weights and approval policy so the selected tradeoff can be reviewed.

Revalidate at commitment time using fresh authoritative state and an idempotency key. Reservations must be atomic or protected by version checks. If another process consumes capacity after planning, the stale proposal should fail cleanly and return to planning rather than forcing an over-allocation.

The constraint encoding guide provides the implementation pattern for predicates, timing rules, evidence, and unknown verdicts. The important point is that a model proposes and a separate control proves the permitted action.

Double allocation occurs when agents read the same available balance and commit independently. Detect it with concurrent scenarios, delayed writes, duplicate messages, and retries. Assert final resource state, not just whether each individual tool call reported success.

Unit and granularity errors make false feasibility look convincing. Cases should mix pallets and units, kilograms and pounds, calendar days and business days, local and UTC times, full-truckload minimums, production yields, and packaging multiples. Every conversion needs an authoritative rule and traceable arithmetic.

Temporal races occur when a prerequisite changes after one agent checks it but before another acts. Test expired quotes, missed dock cutoffs, late quality release, route closures, and a production delay that invalidates a downstream booking. The workflow needs revalidation points, not one approval at the beginning.

Local optimization can create global harm. A procurement agent minimizes unit cost by ordering excess stock. A logistics agent maximizes on-time delivery with unbounded premium freight. An inventory agent protects one site by starving another. Evaluation should compare the combined objective and expose which local reward drove the conflict.

Recovery can consume resources twice. A timed-out tool call may have succeeded, so a blind retry can duplicate a purchase order or booking. Test ambiguous acknowledgements, partial success, and compensating actions. Idempotency and reconciliation are part of correctness.

Constraint blindness is especially costly when agents treat hard limits as preferences. The constraint-blindness analysis explains why repeating a budget or capacity limit in prompts is weaker than verifying it at the action boundary.

Static planning tests are not enough. Introduce a disruption after the system has reserved some resources: a supplier rejects a lot, a port closes, demand rises, a quality hold is issued, or a carrier reports a delay. The system must identify which commitments are affected before proposing recovery.

Preserve commitments that remain valid. Full replanning can create unnecessary churn, duplicate actions, and new risk. Measure plan stability, number of changed commitments, recovery time, incremental cost, and service impact alongside feasibility.

Encode authority by consequence. Routine substitutions inside an approved supplier and cost envelope may be automatic. A new vendor, customer priority change, safety-stock breach, or large premium freight charge may require a named human role. Approval tokens need scope, expiry, and connection to the exact plan version.

Test rejection and delay. A reviewer may deny a proposal, request evidence, approve only part of it, or fail to respond before a cutoff. The agents should not reinterpret silence as approval or reuse approval after the underlying quantities change.

Define one authoritative source and version strategy for every shared resource. Verify units, effective times, reservation states, ownership, and reconciliation behavior. If two systems can update the same balance, test their concurrency explicitly.

Cover normal demand, boundary values, scarcity, infeasible requests, conflicting priorities, stale facts, partial fulfillment, duplicate events, out-of-order messages, tool outages, and malicious supplier content. Include repeated runs so nondeterministic planning paths cannot escape the hard gates.

Assert plan feasibility before execution, revalidation at commit, atomic reservation, delegated authority, idempotent external calls, trace completeness, and safe rollback or reconciliation. Evaluate final real-world state after side effects, not only agent messages.

Score service, total landed cost, resource utilization, plan stability, time to recover, manual intervention, and constraint margin only among feasible plans. Report results by scenario class and root cause so an average does not hide systematic failure under scarcity.

Connect avoided expedites, line stoppages, excess inventory, manual planning, and exception recovery to the ROI framework for agent testing. The strongest business case comes from prevented operational consequences, not the number of scenarios executed.