Build With Agents. Run Without Them.
The first instinct, once you have agents that work, is to put them everywhere.
In every workflow. Every process. Every decision. It feels like progress. It is usually a mistake.
The hard part of agentic AI is not building the agents. That problem is increasingly solved. The hard part is the discipline that comes after: deciding where agents belong, where they do not, and how you keep control of them once there are hundreds of them running across your operation.
Two disciplines separate a demo that impresses a steering committee from a system a regulated institution can actually trust in production. The first is how you govern your agents. The second is how you execute the work they design.
Most of the conversation skips both. We think they are where the real maturity lives.
You cannot scale what you cannot govern centrally
The moment agents become easy to create is the moment they become easy to lose control of.
This is the quiet risk in the agentic era. When every team can spin up an agent to solve a local problem, you do not get an intelligent operation. You get agent sprawl: dozens of autonomous processes, each with its own data access, its own tool permissions, its own escalation logic, none of it visible from a single place. Shadow agents are the new shadow IT. And in a regulated environment, that is not a productivity story. It is an audit failure waiting to happen.
The answer is not to slow down agent creation. It is to govern it centrally.
Centralised agentic governance means there is one place where every agent in your operation is defined, permissioned, monitored, and audited. One place to see which data objects an agent can touch, which tools it is authorised to invoke, what it is allowed to decide on its own, and where it must hand off to a human. One place to answer the question a regulator will eventually ask: what is this agent allowed to do, and who decided that?
Without this layer, every new agent increases your risk surface. With it, every new agent inherits the same controls, the same boundaries, the same accountability. Governance stops being a brake on adoption and becomes the thing that makes adoption safe to accelerate.
You do not govern agents to constrain them. You govern them so you can deploy more of them with confidence.
Build with agents. Run without them.
Here is the second discipline, and it is the one that quietly determines whether your automation is economically viable at scale.
An agent is extraordinary at handling ambiguity. It understands context, absorbs variation, and navigates the hyper-customised reality of investment operations: the firm-specific thresholds, the jurisdictional nuances, the client-by-client exceptions that have always defied rigid software. This is exactly the flexibility our industry needs, and exactly what no static tool has ever delivered.
So separate the two things agents are actually for.
Use the agent at design time: to create, configure, and refine the workflow. Let it ideate the logic, map the exceptions, translate a messy operational requirement into a precise, structured process. This is where its intelligence and flexibility are worth every token.
Then take the agent out of the loop and let the workflow run deterministically. No model call. No probabilistic step. Just a reproducible, auditable, low-cost process executing the logic the agent helped design.
The agent builds the engine. The engine runs without the agent.
This is the difference between intelligence and a capability. An agent that reasons through every execution is a research project. An agent that designs a deterministic workflow you can run a million times for a fraction of the cost is infrastructure.
The strength of the agent is in the ideation and construction of the workflow. The strength of the workflow is in its execution. Confuse the two and you pay an agent tax on every single run. Separate them, and you get the customisation agents make possible with the cost, speed, and traceability that only deterministic execution can deliver.
This is precisely the architecture SPARK is built on. Agents do not replace deterministic workflows. They create them, edit them, update them, and orchestrate them. The workflow engine does the running.
The same principle, twice
These two disciplines are the same idea, applied in two places.
Governance decides where agents are allowed to act. Deterministic execution decides how often they need to. One controls the boundary. The other controls the cost. Both are expressions of the same conviction that has run through everything we have built: the goal is not maximum autonomy. It is the right autonomy.
Agents everywhere, calling models on every execution, ungoverned and unaudited, is not a vision of the future. It is an unsustainable bill and an unmanageable risk.
Agents governed from a single point of control, used to build deterministic workflows that then run on their own, that is an operation that can scale.
The question was never whether agents are powerful. They are.
The question is whether you have the discipline to use them where they belong, and the infrastructure to run everything else without them.
We do.
Next Gate Tech builds SPARK, the operating ecosystem for agentic AI automation of investment operations. Clean data, deterministic workflows, domain knowledge, and full traceability. Purpose-built for one of the most heavily regulated industries in the world. Trusted by BNP Paribas Securities Services, Amundi, Capital Group, Eurizon and others across €500bn+ in assets.