How much may an AI agent decide on its own?
Witold Reichhart · witold.reichhart@capgemini.com
Part one draws the map before any case. It starts from the question clients ask about control, shows why everyone means something else by it, and places their answers on the map, by altitude and essential. Then it places the words engineers already use on that map, lists the parts of each essential with their owners, and settles what the control plane is, a definition the rest of the set relies on.
Read it first: parts two, three, and four assume the map it draws.
About 20 minutes. → or the space bar moves on, and through the steps first where a slide builds in steps. Click a line on the left to go straight to that slide.
M opens the contents, F finds a word, and T switches between light and dark.
Then part two follows one decision through the essentials, part three opens each essential for specialists, and part four turns the map to your firm.
“We have built dozens of agents, on three platforms and with four models. Am I in control of what they are all doing, and am I governing them properly?”
Each of them is right about one part. Nobody can yet say how the parts fit together.
The law sets what applies, supervisors and auditors examine, customers receive the decisions, and third parties run much of the technology. The firm answers for all of it.
The firm answers for every decision taken on its behalf, whether a person, a system, or an agent takes it.
The same work is governed three times: as policy for the whole firm, as authority for one process, and as controls in the systems. Each has its own rhythm, and any of them can change at once when something goes wrong.
An entitlement lets someone reach a system. A mandate says what they may do there.
The three altitudes locate responsibility: the enterprise, the workflow, and the platform. The essentials say what must be governed at each of them: the agent, its knowledge, its mandate, the check, the record, and the loop. Each is a question anyone who acts for the firm has to answer; the first three are settled before the act, the fourth at the act, and the last two after it.
The first three must hold. The fourth checks them. The last two remember, and learn.
Technical people arrive with a vocabulary of their own: the harness, the guardrails, the gateway, the policy engine. Each word lives in one box of the map.
Every essential is a set of parts, and every essential has owners who answer for it before the agent acts. Most of the parts can be bought. A few have to be joined.
Part three opens the essentials one at a time, for specialists.
The answers from the start, placed on the map. Each lights the boxes its owner thinks about.
Each of them is right about their part. The map is every box, and the links between them.
A control plane instructs and can refuse; a management plane watches and reports. Platforms already instruct: they route each call to a model, enforce guardrails, and grant access. For delegated decisions the plane also needs the part that can refuse an act: the mandate and the check, joined to the record and the loop.
What the platforms already do, for their own agents: who the agent is, which model answers, what may go in and out, which systems it may reach.
w3 holds what the agent may decide, as data. w4 and p4 read it before every act, in the path of the act, and permit, refer, or block.
Most of the boxes on the map can be bought or already exist. The three joins that make them a control plane have to be built, on the stack the firm already runs.
Every box on the map is drawn for one agent taking one act under one mandate. Above or beyond that are delegation, cohorts, cascades, other firms’ agents, and the agent’s own foresight: a firm that runs more than a handful of agents meets the first four, and the last is where the agents themselves are being taken.
Whose mandate does the second act under, which check reads it, and does the record of the first name the second? The seventh case shows it: a credit agent with two delegates, under one mandate and its two sub-mandates.
Three elections on one issuer, each permitted, together over the concentration limit: locally correct, globally non-compliant, which the papers call dynamics blindness. The limit is a ceiling at e3, written so the check can read it; the check needs the sum of the permits in flight against it, which no single mandate holds. Situation card 5 plays it for one issuer; the limit no ceiling names in advance is what the map does not draw.
A acts, B reads the changed state, C’s act now conflicts with both. Nothing any one of them did was wrong, and the firm’s position is. The failure sits in the dynamics between acts, where no box looks.
Two control planes, two mandates, two records, and nobody who can see both. What one firm permits is an input the other never declared.
Today’s agents infer the next token. The next ones are being built to carry a world model, like JEPA, and to plan an act by predicting what it will do. The map does not move: the predicted consequence enters the check as one more claim, one statement the agent relies on, with a standing, and the record keeps what the agent expected beside what happened.
No agent sees the whole picture, and none needs to. Each reads what the others left behind: a permit in flight, a state another changed, a claim another relied on. Those traces live on the map already: the record, w5, is where an agent leaves one; the estate, e2, and the knowledge that applies, w2, are the medium the traces share; the current state the check reads, w4, is their sum. So governing a cohort starts with governing the traces: who may leave one, with what standing, and who reads it before the next act.
Regulation governs the constraints and the control plane governs the act. What emerges when many agents coordinate is a territory of its own, and the set signals it here without drawing it: the reference model holds the question, and the seventh case draws the first step, one chain of three.
Part two takes one decision through the map, essential by essential: the agent itself, what it knows, what it may decide, the check at the act, the record, and the loop. Then the demonstration. The case is chosen for the audience; the sixth shows the same map at work outside financial services, and the seventh shows one agent delegating to two.
A motor claim of €3,400
A payment of €18,500, held on a sanctions-screening hit
The renewal of a €180,000 SME credit line
An election on an optional dividend, for 60,000 shares
A repayment offer to a customer in arrears, in a conversation
The renewal of a housing allowance
A €220,000 SME renewal by a credit agent and two delegates