Everyone's racing to build AI agents but almost no one's thinking about what happens when they have more than one. #IBMPartner In this video, I talk about how once you spin up a couple of agents, they start doing real work, and it looks like progress. Then you add a third, and suddenly they're contradicting each other, duplicating effort, and taking actions nobody can trace. Picture an orchestra of world class musicians, each one blindfolded and wearing noise canceling headphones. Every individual is brilliant, but together they're an expensive mess. That's the story of agentic AI at most companies these days. An expensive mess indeed. The fix isn't more agents; it's orchestration. A control plane that defines what each agent does, what data it can touch, how it hands off, and when it escalates to a human. It's the connective tissue that turns scattered capabilities into work that actually gets done end to end. IBM's watsonx Orchestrate is built for exactly this: coordinating agents, workflows, and apps with security and governance baked in from the start. Explore more here: https://ibm.biz/~CSyfa7Bku #sponsored
The orchestra image lands, but in the rollouts I have seen the first thing that breaks is not coordination, it is shared state. Two agents read the same record ten seconds apart, one acts on a stale version, and nobody can tell later which one was right. A control plane only helps if it also owns the write path, not just the routing. The question I now ask before agent three goes live is simple: what is the single place this system reads truth from, and who is allowed to change it?
Funny how even small SaaS teams fall into this, multiple bots and scripts running around without a real handoff plan. Debugging the aftermath usually takes longer than setting up orchestration upfront.
the orchestration layer isn't the fix if the agents weren't scoped tightly enough to begin with. a control plane coordinating three vague agents is just organized chaos with better logs.
This is the problem nobody demos, and it's very real. The orchestra metaphor nails it. Where I'd push a bit: the fix people reach for first is routing, who does what and who hands off to whom. That part is relatively easy. The reason two agents actually clash is dumber than it sounds: each has its own memory and can't see what the other already did. So one emails the customer, the other emails them again an hour later, both convinced they were being helpful. What saved us was boring: one source of truth every agent reads and writes to, plus a log of who did what and why, so a human can retrace it when it goes sideways. Routing is what gets demoed. Whether the thing survives production comes down to the shared state and the trace, and that never makes the keynote.
Spot on. Orchestration is the real bottleneck once you scale beyond a single agent. We've found that defining handoffs and data access is just step one. You also need a verification layer to reconcile it when agents inevitably return conflicting or overlapping information. Without that traceability, the mess just gets expensive faster.
Three agents optimizing correctly in isolation can still produce contradictory outputs. That's the failure mode orchestration pitches never address. Skill isn't the gap, shared context is. Most teams still measure success per agent, not per outcome, and that's exactly where coordination breaks down.
The orchestra metaphor undersells the real risk. In regulated industries the failure mode isn't just duplication or contradiction -- it's unauditable decision chains. When agent three overrides agent one's output in a credit or compliance workflow, you need a full provenance trail or you're facing regulatory exposure. Orchestration without lineage is just organized chaos.
Agent sprawl is becoming the new application sprawl. The challenge isn't building more agents, it's governing how they collaborate, share context, and make decisions.
The orchestra analogy is spot on, and from a data engineering perspective, the problem runs deeper than coordination. When agents share access to pipelines, databases and APIs without a clear control plane, you lose lineage too. You can't tell which agent read what, when, or why a downstream decision came out the way it did. That's a governance headache before it's even a product problem. The orchestration layer isn't just operational glue; it's also your audit trail.
Cross-Border Architect | Founder of GloireCompany | International Newsletter Author and Editorial Strategist | Building Long-Term Opportunities Across Borders | Creator of the Thirty-Five Axis Extension Framework
1wAn excellent analogy, Cassie. What resonates with me is that the challenge is no longer building intelligent AI agents. It is designing intelligent systems around them. Without clear orchestration, governance and accountability, adding more capable agents can simply multiply complexity rather than productivity. I also appreciate your emphasis on the control plane. As organisations adopt agentic AI at scale, defining responsibilities, data access, decision boundaries and human oversight will become just as important as model performance itself. For me, the future of enterprise AI is not determined by how many agents an organisation deploys. It is determined by how effectively those agents collaborate within a well-designed operational framework. Thank you for sharing these valuable insights. As AI ecosystems continue to evolve, orchestration will be the foundation that transforms individual capabilities into trusted, measurable business outcomes. Cassie Kozyrkov