Agentic Terminal Operating System
A terminal operating system where containers, equipment, and carrier services actively participate in planning and coordination.
PortSim.ai treats the port not as ten thousand moves to schedule, but as thousands of overlapping intentions to reconcile—continuously, and in everyone’s favor.
A large terminal isn’t really managing ten thousand container moves a day. It is continuously reconciling thousands of overlapping intentions about those containers and the resources around them.
A single box can sit inside many intentions at once: the cargo owner wants it in Detroit by Tuesday under budget, the line wants it off the vessel in its discharge window, customs won’t release it until it clears, the railroad is building a block at 22:00, and the yard wants to stay below 80% utilization. None of these, alone, says what should happen next. What happens next emerges from their interaction.
That is why intention is a far more powerful abstraction for agentic AI than “task,” “command,” or even “goal.” PortSim.ai is built to model it directly.
One box. Many parties. One identity.
“The container should be in Detroit” is a goal. “I intend this container delivered to Detroit by 14:00 Tuesday, by rail if possible, under $1,200, pending customs clearance” is something an agent can actually reason about.
Intentions don’t need to “sum” to anything. There simply needs to exist a future trajectory that falls inside the acceptable region of every sufficiently authoritative intention. The system’s job is to find that intersection—and keep it alive as the world changes.
And when there is no intersection, the agent doesn’t hallucinate a “best” route. It reports that the intention set has become infeasible and asks which intention may be relaxed—the deadline, the budget, or the routing.
A multi-agent terminal operating system
Containers, equipment, and carrier services stop being things to be scheduled and become participants that plan, offer, commit, and negotiate—coordinated by a terminal that balances every delivery against the health of the whole operation.
Each container carries a software agent that represents its delivery obligations, budget, and handling requirements. It watches the box’s situation, spots threats to its journey, weighs alternatives, and negotiates for the handling and transport it needs.
Agents for cranes, straddle carriers, yard blocks, reefer plugs, storage, and carrier services offer capacity and make firm commitments—but only within the authority they’ve been delegated.
A coordinator brings the individual plans together—resolving competition for shared resources and balancing each delivery commitment against the performance of the terminal as a whole.
Not every intention is commensurable, so the system never trades a customs hold for a few dollars of efficiency. Law and safety sit above regulation, custody, contracts, and operations—encoding who may intend what, about which objects, under which circumstances.
When equipment fails or circumstances shift, affected agents revise their plans and renegotiate their commitments. When an intention set becomes impossible, the system surfaces the conflict instead of quietly failing.
Underneath it all is a reproducible, deterministic discrete-event model of the terminal. Different controllers face the same rules, workloads, and disruptions—and see only the information they’re authorized to—so better decisions can be told apart from better luck.
At scale, the terminal becomes a living graph. Edges aren’t just relationships—they’re intentions, commitments, dependencies, prohibitions, and reservations. Logistics optimization becomes continuous intention reconciliation over a changing world.
To send Container A to rail, the plan has to secure Container C’s destination, a rehandle, a transfer vehicle, and rail acceptance—then execute in physical order: rehandle C, pick up A, move A, load A. A dependency, not a preference.
Four commitments shape everything we build.
We model what each actor is trying to bring about—not just the next command to run. Outcomes emerge from the interaction of intentions, and the system reasons about that directly.
Safety and law aren’t line items in a weighted score. Precedence and the right to intend are first-class—the system respects who is entitled to decide what.
Decision-making is kept separate from the world’s rules. Every approach faces the same operation, and runs replay exactly—so a claimed gain is a real gain, not a difference in the test.
Operators keep control over decisions that exceed the system’s authority—raising a budget, changing a delivery promise. The AI maintains the picture; people make the calls that are theirs to make.
We measure the ground before we claim it.
The simulator is a controlled environment for honest comparison. Decisions, events, and results are logged so any run can be reproduced, replayed, and investigated—and so a smarter controller can be told apart from one that simply saw more.
~1,300
tests across the platform, plus checks against established queueing models.
80
replication comparison of conventional scheduling approaches, honestly reported.
30
documented gaps between the model and real terminal operations, tracked openly.
Replay
deterministic runs, reproducible from event logs down to the decision.
In our baseline study, a more complex planner showed no clear advantage over a simple greedy method in two scenarios—and did worse in the largest. That kind of finding is exactly why the benchmark comes first.
A foundation you can trust, then agents that use it
The rule is simple: agents change how decisions are made. They never change the rules of the simulated world.
An operating model of the terminal with conventional dispatching, real resource and process constraints, stacks and rehandles, vessel, rail and gate flows, schematic animation, scenario tools, replay, and validation evidence. It is useful on its own—and it establishes the controller interface the agents will later plug into.
Delegated container missions, feasible-option reasoning, negotiation and firm resource commitments, authority and fairness, failure recovery, and human intervention—all compared head-to-head against conventional dispatch on the very same operation.
First use case
Making a container’s connection is real work: a rehandle to dig the target out of the stack, a CHE cycle, internal transfer, release (COREOR / customs hold), and the handoff inside the rail cutoff or gate appointment window. PortSim.ai coordinates all of it—then tests whether delivery performance actually improves over conventional dispatch, and whether those gains hold across the terminal instead of creating delays somewhere else.
Run a terminal, a carrier, or a logistics platform—or just as interested in intention-based agents as we are? See plans, or call us.
PortSim.ai
A venture of The Infogetics Group, LLC