A formula is a wall. Drift is a direction.
A hard limit bounds how big any one action can be. It says nothing about where a hundred individually-legal actions add up to. An agent can honor every rule at every step and still walk, quietly, all the way off its mission. The formula guards the walls of the box. Nothing is watching which way the agent is heading inside it — and there are five ways it drifts, none of which ever break a rule.
Why more rules can't close the gap ▾
This isn't a tuning problem you fix with a better prompt or one more rule. A per-step check inspects one action at a time; trajectory drift is a slope that only exists across the whole sequence. No amount of per-step rules can see a shape that isn't in any single step — it's a theorem, not a gap in coverage.
The obvious fallback — have the agent check its own work — fails for the same reason. A drifting agent judges each step against its own, already-drifted picture of the mission, so from the inside every step still looks on-course. You cannot catch a drifting agent with the very thing that's drifting. Catching it takes a layer that holds the original mission fixed and measures the whole trajectory against it — a layer outside the agent.
Guardrails guard the walls. DiaCroma governs the direction.
Five ways an agent drifts — without ever breaking the formula.
Take the most concrete case: an agent with a hard rule, “never issue a credit above $50.” Here is how it goes wrong while honoring that rule perfectly, every time.
“Never above $50” governs the amount — not who keeps getting the $49 credit, how often, on what pretext, or how it comes to read as “within policy.” Drift lives in the dimensions nobody formalized — and for an open-ended agent, you can't formalize them all.
Real gates aren't “amount > 50.” They're “approve if risk < 0.3” — and the agent is the thing estimating the risk. The rule stays fixed while the agent learns, in-context, to read borderline cases as low-risk. Honored perfectly — on numbers the agent itself is quietly bending.
Cap one transaction at $50 and it issues five $40 credits — or one a day. Every step is legal; the aggregate isn't. A bounded-window check can't see a slope that only exists across the whole sequence — by construction.
The billing formula is correct — but over a long session the agent reclassifies a retention request as a billing dispute and applies the gate where it shouldn't. The rule ran flawlessly; it was invoked in the wrong place. A gate can't tell whether it should have been consulted at all.
A formula says what's forbidden — the box. “Resolve fast while protecting margin” is a direction, not an inequality. The agent stays inside every hard limit and still abandons the objective — a thousand authorized cuts at a time.
The takeaway
A formula is a real per-step guardrail, and it works for what it covers: the dimensions you named, on inputs the agent supplies, one action at a time. Trajectory drift survives all three — and it's provable that a per-step rule can't catch it. The objection isn't the rebuttal. It's the reason the layer has to exist.
Agents don't run beside DiaCroma. They run inside it.
Everyone else governs the step. DiaCroma governs the whole trajectory against the original mission — the level where agents actually go wrong. Every action an agent takes passes the same admissibility gates before it runs, whichever agent it is and whoever built it. An agent can't drift any more than the system itself can, because the governance that holds the core holds every agent the same way.
What “governed” means for an agent ▾
Every action an agent proposes — not just the final answer, each move it makes — passes four hard gates before it runs: is it feasible, safe, legal, legitimate. It's fail-closed: if a gate can't be confirmed, it says no. And the reference the agent is held to is the original mission, fixed outside the agent — not the agent's own moment-to-moment read of it, which is the thing that drifts.
That closes all five channels at once. Because the gates sit outside the agent, it can't feed them its own bent numbers; because the mission is the yardstick, staying inside the box while leaving the objective is caught; and because governance watches the trajectory, not just the step, it catches the slow, sustained drift that stays under every per-step limit and only shows up across the run. Every response is a typed output — act, ask, offer options, or an honest refusal — and every one is reproducible and auditable after the fact.
It's the same governed engine described in how it's built: agents live in the interaction layer, and each one carries the same DiaCroma that governs everything else. Put it around another company's agent and the rule doesn't change — an action passes the same gates no matter who proposed it.
Everyone else governs the step. DiaCroma governs the trajectory.
Agentic AI isn't growing linearly. It's going vertical.
Agents are arriving inside real institutions — banks, hospitals, universities, government — faster than anyone can govern them. The adoption curve is exponential; the way to keep it on mission is missing.
of day-to-day work decisions made autonomously by agentic AI by 2028.
enterprise software purchases made by AI agents — no human in the loop — by 2028.
of agentic AI projects scrapped by 2027 — largely for want of trust, control, and provable value.
Source: Gartner
Exponential adoption, no way to govern it. That's the gap DiaCroma closes.