Lab
What changed in agentic AI, explained without jargon.
Five recent shifts, what they mean for an airport that has to prove its work, and what we do about each of them.
The model is not the product. The harness is.
What it is. A language model on its own answers questions. An agent is the model plus everything around it: the tools it may call, the rules that stop it, the budget it may spend, the memory it keeps, the checks its output must pass. That surrounding structure is the harness.
Why it matters to you. Two teams with the same model get very different agents. Reliability, safety and cost come from the harness — and so does accountability. It is also what lets the model be replaced when a better one ships, without rebuilding the operation.
What we do. Every Khelion agent runs in the same harness: warrant enforced in code, budget per task, anti-loop, chained ledger, tests on outputs. What you buy is the harness; the model is a component.
Several narrow agents beat one that does everything.
What it is. Instead of one assistant asked to do a bit of everything, a team of small agents, each with a named role, coordinated by another agent that routes work, tracks dependencies and escalates. They wake up when work arrives — a new file, a date, a request — not on a timer.
Why it matters to you. Narrow roles are easier to specify, to test and to limit. When one is wrong, you know which, and why. The output is files in your systems, which your people can open and correct.
What we do. Small teams across six airport areas — airside, security, baggage, passengers, safety, environment — a coordinator, and a place for your own process. Deliverables land where your teams already work.
Not every task deserves the largest model.
What it is. Model providers now offer several sizes of model and, on the same model, several levels of “effort” — how long it reasons before answering. Repeated context can be cached instead of being paid for again. Together these make cost a design variable rather than a bill you discover afterwards.
Why it matters to you. A drafting task, a classification and an architecture decision do not need the same intelligence. Sending everything to the most capable model is how agent programmes become expensive and slow without becoming better.
What we do. Each task class is routed to the smallest model and the least effort that passes its check; the most capable models are reserved for architecture and governance decisions. Every run records what it consumed, so the routing is measured, not guessed.
Limits enforced by code, not by asking the model nicely.
What it is. An agent can be talked into things — by a document it reads, by an email, by a web page. Prompt injection is the name for it. The answer is not a better prompt; it is a structure where the model cannot execute what the warrant forbids, whatever it was told.
Why it matters to you. Your agents will read untrusted content: supplier emails, PDFs, tickets. The question an auditor asks is not “does the model behave?” but “what happens when it doesn’t?”
What we do. Least privilege on every connection; secrets never pass through a prompt; write access only where the warrant allows it; content read by an agent is treated as data, never as instructions; every refusal and stop lands in the ledger. Processing in the EU by default, or inside your environment; no client data used to train anything.
The cost of an agent is tokens, times attempts, times the people re-checking it.
What it is. The visible cost of an agent is what the model provider bills. The real cost adds the runs that fail and are retried, and the human hours spent checking work that should have been right the first time.
Why it matters to you. An agent that produces reports nobody reads, or that needs a person to redo half the work, is a cost, not an asset — whatever the demo looked like.
What we do. One measure agreed before day one; deliverables instead of reports; a budget per task; runs that fail are counted and fixed; and a stop rule that we apply to ourselves first. Our own operation runs on the same rules, which is where most of these lessons came from.