Soon we will have thousands, maybe millions of agents in every enterprise. At this point the control plane for an agent fleet, or agent ecosystem, will be the next clear and present enterprise architecture challenge. Kubernetes, microservices, identity, security, and observability, all at-scale - are now the required operating model for enterprise agents.

1/ Enterprise agents are not personal agents.

We draw a clear line between laptop-bound agents that run under a user’s identity and enterprise agents that must be secure, observable, operable, reliable, and manageable like any other enterprise application.

2/ Microservices become micro-agents.

Our core argument is that agents should be packaged as smart microservices in containers, which lets enterprises use proven cloud-native patterns for deployment, scaling, security, and operations rather than inventing an entirely new runtime model.

3/ Fleets require a real control plane.

We frame the challenge as managing thousands or even millions of agents, organized into observer, goal-oriented, and task-oriented roles, with Kubernetes serving as the deployment and coordination substrate for fleets, and even fleets of fleets.

4/ The control plane for agent fleets is therefore much broader than orchestration alone.

It is about Kubernetes, microservices, security, identity, state, communication fabric, trust, and the surrounding services that make agents usable in production.

5/ This is an important and urgent challenge that we need to address!

We believe that the next phase of enterprise agents will depend less on whether a single agent can do something impressive, and more on whether a fleet can be governed, trusted, and run as part of the operating model.

Special thanks to:
Cloud Native Community Groups https://community.cncf.io/
The Cloud Native Computing Foundation @cncf
Josh Halley, Principal Architect at Cisco Systems
and to all of the virtual participants.

Looking for more?
👉 Discover the full O’Reilly Agentic Mesh book by Eric Broda and Davis Broda

🎧 Follow co-hosts John Miller and Eric Broda on The Agentic Mesh Podcast on Youtube, Spotify and Apple Podcasts. A new video every week!

Agentic Mesh Webinar with Actian

(yes, I can proudly say my son is the co-author of the book!) and I had a great conversation with at Actian () about our new O’Reilly book, Agentic Mesh, and what it takes to run agents at enterprise scale. There were many topics covered, but here are a few of the highlights:

1/ Why a mesh, not a monolith

Industry leaders — Jassy (CEO, Amazon), Nadella(CEO, Microsoft), Huang (CEO, NVIDIA) — project thousands to millions of agents per enterprise. At that volume, a monolithic architecture is not deployable; a federated mesh where any agent can collaborate with any other is the only practical path.

2/ Enterprise-grade agents are not chatbots

We define seven characteristics that separate production agents from prototypes: security, reliability, explainability, scalability, discoverability, observability, and operability. These are the same disciplines applied to any critical enterprise application — applied now to agents.

3/ Testing agents requires a different approach

Traditional equality-based software testing breaks down when LLM outputs vary across identical inputs. We think a better approach is to use evaluation rubrics indexed by autonomy and criticality, and judge agents that assess intent rather than exact output.

4/ The knowledge layer is the next bottleneck

Getting the right context from petabytes of enterprise data into a token-budget-constrained agent window is the core scaling problem. Taxonomies, ontologies, and knowledge graphs — what we call the Agentic Knowledge Fabric — sit at the center of that solution.

Looking for more?
👉 Discover the full O’Reilly Agentic Mesh book by Eric Broda and Davis Broda

🎧 Follow co-hosts John Miller and Eric Broda on The Agentic Mesh Podcast on Youtube, Spotify and Apple Podcasts. A new video every week!