Nishant Gupta, Staff Software Engineer & Tech Lead at Meta, joins us in the studio at Agent Conf 2026 in Warsaw to discuss AI agent fleets and the infrastructure needed to run them.
We talk about how agents share responsibilities, communicate, and work toward a common objective. Nishant describes engineering and operations use cases at Meta, explains why authentication and observability belong in a reusable control plane, and shares what it would take for him to trust agents with more complex work.
What makes a group of agents a fleet?
Nishant views agent fleets and agent swarms as the same idea: agents with distinct responsibilities, communicating and modifying shared state to complete a task.
He uses refunds and event logistics as examples. One agent might identify an anomaly, another might contact the relevant team, and a third might act on the refund. The question is, how do those responsibilities fit together?
Agent fleets for code quality and operations
Nishant describes an engineering workflow in which agents analyze code quality, test simplifications in sandbox environments, update dashboards, and submit pull requests for senior engineers to review. Other agents help explain changes to the architecture and update on-call runbooks.
He also shares an operations example at Meta, where agents run in the background across products and services, examining past incidents and the current codebase, running simulations and failure injection tests, and surfacing potential root causes to engineers. Some can be configured to make fixes, with a person still reviewing the result.
A reusable control plane for changing agents
As agents and frameworks evolve, teams need infrastructure they can reuse. We ask Nishant about the control plane around a fleet, including observability and cost ceilings.
He explains why repeatedly building authentication, tracing, and stable infrastructure is expensive. His approach is to keep that shared layer relatively stable while the agents underneath it change, allowing teams to reuse the infrastructure across different fleets.
Where to start building an agent fleet
We also ask what tools developers can use as they move from a few manually configured agents to larger fleets. Nishant points to the agent tooling and documentation available from Google, AWS, and Azure, including capabilities for identity and observability. He suggests choosing a provider based on the team's familiarity and how quickly it can learn the tools.
Trust takes repeated, reliable outcomes
Nishant says that verifying agent outcomes still takes a lot of time, and defining requirements and giving an agent tools does not remove the need to check its work.
For him, confidence would come from seeing agents solve problems reliably over months. That consistency would let him delegate increasingly complex tasks. We close by discussing what that could mean for engineering productivity.
About the guest
Nishant Gupta is a Staff Software Engineer & Tech Lead at Meta, working on AI infrastructure and distributed systems. At Agent Conf 2026, he presented "The Agent Control Plane: Managing Thousands of AI Agents Without Losing Control." This studio interview follows up on the architecture, use cases, and trust questions behind agent fleets.
Resources
