Orchestrate, Accelerate and Govern: Why Enterprises Need an Agentic Control Plane
At ALIGN Chicago 2026, IBM watsonx Orchestrate director Gargi Dasgupta argued that agent sprawl is an operating model problem and demonstrated how a control plane detects, fixes, evaluates and governs agents in production.
By The AI Enterprise Magazine Team3 min read
Key takeaways
- AI has reached the core. Agents now touch procurement, finance and HR, not just customer care.
- Sprawl is the new risk. Easy agent building has scaled fragmentation.
- The gap is the operating model. Standards exist for connecting agents, but not for governing them.
- Building is 20 percent. Integration, evaluation, monitoring and access control are the rest.
- Agents govern agents. A control plane agent can flag problems across thousands of agents.
Building an AI agent now takes days. Running hundreds or thousands of them safely is the hard part. Gargi Dasgupta, Director of IBM watsonx Orchestrate, argued that the next era of enterprise AI spending will go to control: the evaluation, monitoring, security and governance that let agents run in core business processes. Dasgupta spoke at the ALIGN AI Executive Summit Chicago 2026 on July 15, in a keynote sponsored by IBM titled Orchestrate, Accelerate and Govern the Agentic Enterprise.
What has changed about where AI runs?
A few years ago, Dasgupta said, AI sat at the edges of the enterprise, in customer care and experiments. Now it reaches procurement, finance and HR. In IBM's conversations with C-suite clients, she said, about 70 percent say AI is already touching their core business.
The technology path ran from predictive machine learning, judged on precision and recall, to generative models that made assistants contextual and multi-turn, to agents that plan, act through enterprise APIs and learn from the results. "It's not a hype. The technology will only get better," she said.
Why is agent sprawl a problem?
Because building agents is now easy, every hyperscaler, startup and software vendor offers an agent builder, and teams across the enterprise are using them. "We didn't scale AI; we scaled fragmentation," Dasgupta said. Standards such as MCP for connecting agents to enterprise APIs and agent-to-agent protocols for letting them talk have helped, but there is no comparable standard for governance.
The underlying problem, she said, is not the technology but an immature operating model. That model has several layers: hybrid infrastructure that lets companies run models where they choose, with sovereignty a priority for many clients; data that is still siloed; and an agentic control plane that governs the agents themselves.
Where does the real work go?
IBM's experience over the past year is that building an agent is about 20 percent of the work, Dasgupta said. The other 80 percent is integrating with back-end systems, evaluating probabilistic systems that can't be tested with simple pass/fail checks, monitoring agents in production when conditions change, and controlling access so agents get "just enough and just in time." Agents will increasingly be needed to evaluate other agents.
What does a control plane look like in practice?
Dasgupta showed a demo using watsonx Orchestrate and IBM's coding agent, Bob. A telecom customer asks a website agent about a phone recall announced that morning, and the agent, built before the recall, can't help. The customer leaves negative feedback.
An operator responsible for dozens of agents gets a diagnostic alert from a control plane agent that watches all the others, pointing her to the trending complaints. Without that insight, she said, the problem might have run for days or the agent might have been taken offline. Instead, a builder asks Bob to research the issue. Bob finds the recall, proposes a new subagent to handle recall questions and checks with the builder at each step.
Before deployment, enterprise guidelines are applied: security credentials for the CRM connection, a PII filter because the subagent collects contact details, and a fallback model policy. The operator reviews evaluations, replays full interactions to see each step and tool call, confirms the agent meets company policy and deploys it.
What should leaders look for?
Dasgupta summarized three questions a control plane should answer:
- Visibility: How many agents are deployed, how they are used, and their success rates, failures and latency.
- Reliability: Which evaluations have run, whether they are sufficient and whether any agents are in production without them.
- Security: Which agents, tools and models are in use, what guardrails and policies are enforced, and how that matches the company's risk appetite.
The way to act on agentic AI with reliability and safety, she said, is a control plane that is itself powered by agents to govern agents.
Where to hear more in 2027
Enterprise leaders will keep comparing notes on governing agents at the 2027 ALIGN AI Executive Summits, invitation-only gatherings in eight U.S. cities: Atlanta, New York City, Dallas, Chicago, San Francisco, Boston, DC and Los Angeles. Registration is coming soon. See the ALIGN AI Executive Summits page.
Frequently asked questions
What is an agentic control plane?
A layer that gives visibility, evaluation, monitoring, security and policy enforcement across all of an enterprise's agents, regardless of how they were built.
Why is agent sprawl a problem?
Agents built on many platforms across teams create fragmentation without shared governance standards.
What takes most of the effort in deploying agents?
Dasgupta said about 80 percent goes to integration, evaluation, production monitoring and access control.
This article is based on remarks by Gargi Dasgupta, Director, watsonx Orchestrate, IBM at ALIGN AI Executive Summit Chicago 2026 on July 15, 2026. More coverage: ALIGN AI Executive Summits.
Filed under: Agentic AI, AI Governance & Trust, Observability & Data Quality
