By Dan Jaye, CEO 

McKinsey has painted a bold picture of the enterprise future: a dynamic “agentic mesh” where autonomous AI agents collaborate across systems, departments, and workflows to drive efficiency and innovation. If you’ve seen the diagrams, it feels like something out of a sci-fi operating manual – intelligent agents everywhere, working in harmony.

But as Nate’s Substack smartly observed in the piece Software 3.0 vs AI Agentic Mesh: Why McKinsey Got It Wrong, the mesh is mostly marketing. In practice, it’s not just aspirational – it’s largely unworkable in the short term.

Still, before we toss the idea into the same bin as blockchain-for-everything or metaverse meetings, I’d argue there’s something worth salvaging. The agentic mesh is a mess today, but there are valuable capabilities it promised which are still achievable. The problem is that most enterprises and agentic frameworks are nowhere near ready for it.

Let’s take a clear-eyed look at what went wrong, what still makes sense, and what actually needs to happen next.

Why the Mesh Doesn’t Mesh (Yet)

It assumes a level of data maturity that doesn’t exist. McKinsey’s vision relies on real-time, high-fidelity data, stitched together across silos. But most enterprises are still wrangling incomplete data, unstandardized schemas, and pipelines that lag hours – or worse, days – behind. Nate rightly pointed out that the infrastructure we have isn’t built for agent coordination. It’s barely built for reliable reporting.

It skips past identity and context. Agents can’t act without understanding who they’re acting for. That requires durable identity resolution: a system to unify users across devices, touchpoints, and systems. Without that, agents are just guessing.

It decentralizes decision-making without shared oversight. Nate’s post nailed this: distributing control across agents without embedding governance is asking for trouble. Who’s accountable when two agents interpret a goal differently – or worse, act on bad data?

It multiplies complexity faster than it creates value. Each agent added to the mesh brings exponential complexity. McKinsey’s models made it look elegant, but most enterprise workflows are anything but. The more agents you deploy without architectural guardrails, the faster things spiral.

Where the Vision Still Holds Up

Now, before we pile on McKinsey too hard, let’s acknowledge this: the core idea isn’t wrong.

Mesh is the wrong topology.  We need adaptive clusters. The vision of agents coordinating across business domains is directionally correct. But similar to why mesh topologies in graph theory become unworkable and needlessly complex, we want flexible associations for agents that develop strong relationships because of reinforcement from positive outcomes.

“Software 3.0” is a better entry point. Nate proposes something much more realistic: narrowly scoped agents operating in structured workflows, with clear rules and human-in-the-loop oversight. These Software 3.0 agents are easier to deploy, safer to manage, and far more likely to deliver ROI in the near term.

Combine purpose-built siloed agents with practical, curated collaboration. The promise of the McKinsey vision is a continuously improving ecosystem of AI capability.This requires more than parallel development of efficient isolated agents.

This is the way

An autonomous mesh is not the solution, but we need patterns and resources that enable cooperation between agentic solutions.

Start with data readiness – not agent ambition. Before you introduce agents, fix your data: unify it, structure it, clean it, and govern it. If your agents are pulling from dirty, fragmented sources, they’ll produce chaotic outcomes – no matter how advanced they are.

Invest in durable identity. Whether you’re serving customers, managing internal systems, or orchestrating workflows, identity is the common thread. Without it, context collapses. And your mesh turns to mush.

Move one domain at a time. You don’t need a mesh to get started. Pick one functional area – customer service, finance, campaign operations – and build something that works there first. Test it. Govern it. Scale only when ready.

Establish a common language. When multi-vendor, heterogenous agents are stitched together, inconsistent context will be the bane of reliability and utility. Translation losses will just be the start, but when the inherent non-deterministic nature of LLM’s is combined with chaotic semantics, unpredictable outcomes will reign.

Last Thought

Nate’s takedown of the mesh struck a nerve because it spoke the quiet part out loud: most companies aren’t ready. But that doesn’t mean a fluid and dynamic agentic future is nonsense. It just means we need to focus less on diagrams – and more on the plumbing.

The real future of AI in the enterprise won’t be defined by self-directing agents hammering on a million keyboards instead of monkeys, but by reliable foundations for collaboration.  Autonomous agents with exponential relationships are not the answer. However,  in case I’m wrong, let me be the first to welcome our robot overlords.