Many teams we work with are building AI agents one at a time. Each agent ends up with its own:
- login and service account;
- prompt library;
- model, pinned to whatever was best that month;
- blind spots.
That works well for the first few. As the number grows, it becomes hard to say which agent touched which record, under whose authority, using which model.
The agents themselves are rarely the issue. What is often missing is a shared place for them to run.
What "where they live" means
An agent platform needs only a few shared layers, but they must be shared.
1. One identity and policy layer
Every agent signs in as itself and is granted permissions the same way a person is. It is logged against the same audit trail. Being able to answer "who approved this agent to do that?" is what turns a collection of scripts into a platform.
2. A governed model gateway
Models change every few months, and they will keep changing. Put a gateway between agents and models. It routes calls, enforces data rules (what may leave your boundary, what must stay on private models) and records usage and cost. Swapping a model should then be a configuration change, not a rewrite.
3. Shared tools and data access
Agents should call the same governed connectors to your CRM, ERP and document stores. Write an integration once, review its security once, and let every agent use it.
4. Shared evaluation
Every agent is tested against an agreed set of questions before release, refusals included, and again whenever a model changes underneath it.
The payoff
When those layers exist, a new agent stops being a project and becomes a feature. It takes days, not months, and it is safe from the first run because it inherits identity, policy, audit and evaluation instead of reinventing them.
That's the approach we took with yOGI, our enterprise AI platform: one launcher, shared identity, audit and policy across every app, and models behind a governed gateway, in the cloud or fully on-premises.
A quick test for your organisation
- Can you list every agent in production, and who owns it?
- Could you switch one agent's model tomorrow without touching its code?
- Can you show an auditor every action an agent took last month?
If any answer is "not yet", it may be worth investing in the platform before adding the next agent.
Parvesh Kaushal is CTO at 2SD Technologies.
Top comments (0)