As agentic AI shifts from testing to real-world use, companies now face the challenge of managing whole ecosystems of agents, not just individual ones. On a recent Futurise episode, Eric Broda, President and Co-Founder of Broda Group Software, shared his perspective based on years in financial services, systems architecture, and large-scale platform design. Eric’s framing is very clear. Agentic AI is not about individual agents acting in isolation. In fact, it is about fleets of autonomous, LLM-powered agents collaborating, transacting and operating within a broader enterprise ecosystem. He describes it as an agentic mesh. This shift fundamentally changes how organisations must think about governance, trust and operational control.
From Governance to Certification
A defining characteristic of agents is that they are stateful and conversational. Unlike traditional microservices, agents engage in long-running interactions, maintain context, and can take action across systems. This makes conventional governance approaches insufficient. Eric suggests that instead of focusing only on governance, companies should also certify their agents. Like product safety certifications, this approach sets clear purposes, policies, and checks. Agents are certified for specific roles, with set limits and clear trust guidelines. Certification is not a one-time event. In regulated industries, consistency and repeatability matter more than unchecked adaptation. Agents may evolve, but only through controlled, observable processes that preserve compliance with organisational and regulatory requirements.
Trust as a Continuous Process
In this model, trust has several layers. It starts with security, including identity, access control, and encryption, and also covers transparency, explainability, and accountability. Agents must be easy to observe, audit, and diagnose. If their behavior changes, organizations need to know the reasons and timing. Importantly, people—not the agents—are accountable. Each agent has an owner who manages its permissions, limits, and certification. This is similar to how companies already handle applications and users. It also shows that agentic systems should fit into current operations, not avoid them.
Architecture Matters: From Experiments to Meshes
A recurring theme in this exciting discussion is architectural maturity. Many current agent implementations remain confined to single-process experiments. They have impressive demonstrations, but they are unsuitable for scale. Eric argues that achieving agentic mesh requires embracing architectures enterprises already trust. In practice, this means treating each agent as a microservice, packaging it into containers, deploying it via orchestration platforms and securing it using established enterprise patterns. This approach enables scalability, interoperability and integration with existing systems — from identity providers to core transactional platforms. Agent interoperability, while still evolving, is advancing rapidly through emerging standards and protocols. In this model, agents coordinate tasks, but tools execute work. This separation allows organisations to integrate agents across heterogeneous environments without locking themselves into a single vendor ecosystem.
Building for the Long Term
The promise of agentic AI lies not in short-term productivity gains, but in sustained, trustworthy operation over the course of decades. Enterprises must design for observability, certification, interoperability and control from day one. The goal is not merely to build agent teams, but to establish an agentic mesh that boards, regulators and users can trust. As Eric emphasises, trust is not a feature. It is an operating principle. He points out that trust is not just a feature. It is a core principle that will decide if agentic AI becomes a lasting part of enterprise operations or stays limited to experiments.
