where your chatbot dies, AI agent wearing our layer of clothes lives and runs the race.
A chatbot answers and stops. An agent has a goal and keeps going. That difference is the whole reason this layer exists — and the page below explains it.
The think–act–observe loop.
An agent runs a loop: it thinks about the goal, acts on the world, observes what came back, and goes around again. Every pass through that loop is another chance to say something it shouldn't — to a counterparty who remembers every pass.
A chatbot thinks once. An agent plans against a goal — which means it holds context worth extracting: the budget, the deadline, the fallback, the floor.
A chatbot talks to you. An agent talks to the world — suppliers, auditors, regulators, prospects — and each of them has an interest in what it knows.
A chatbot forgets the room. An agent reads the reply and goes around again. Pressure applied on turn four still works, because the agent is still there.
Every dressed agent stands in three relationships at once.
Delegation
You hand it the goal and everything needed to pursue it — the floor, the BATNA, the file. This relationship must be total. An agent you brief partially negotiates blind. The clothes exist precisely so you never have to starve your own agent of context.
Disclosure
What the agent says to the counterparty, per statement, per turn. This relationship must be selective. This is where the clothes live: what is revealed versus concealed, and a different face for the supplier, the auditor, the regulator.
Reading the garment
What the counterparty can infer from everything your agent has said — assembled across turns, the way a buyer reconstructs a floor price. This relationship must be assumed hostile. Not because people are, but because incentives are.
Existing tooling defends the loop from attacks. Nothing defends the loop from its own helpfulness. That gap — statements harmless alone, lethal assembled — is the layer we sew.