The Agent Control Plane: Managing Thousands of AI Agents

Gupta says building a single agent is relatively straightforward; the challenge is tracking what data thousands of agents touch, what access they have, and who is accountable. He groups failures into system design problems, conflicts between agents, and a lack of verification. Examples include agents looping, failing to terminate, or disagreeing over which agent has authority.

Gupta defines the agentic control plane as a governing layer separate from the agents, with seven core functions: life cycle management, policy enforcement, permissions, evaluations, coordination, observability, and human escalation. A unified inventory helps teams find agents and establish ownership. Runtime policies can set token or compute limits, while scoped identities apply least privilege.

Gupta recommends both offline evaluations against a golden dataset and online evaluations using sampled real traffic, with humans or LLMs as judges. Multi-agent coordination needs explicit protocols rather than informal English. Full-fidelity OpenTelemetry traces can record model calls, tool calls, data exchanges, and permissions; predefined escalation triggers can send cases such as refunds above $5,000 to a human.

Gupta recommends gradually routing traffic to new agent versions and comparing them against production on measures such as cost, latency, and throughput. Scoped kill switches can disable a particular agent, tool, or MCP server for selected tenants within seconds. Fleet-wide cost limits and anomaly detection can catch runaway workflows, cost or latency spikes, and drops in evaluation accuracy.

Gupta's roadmap starts with one clearly scoped agent, human approval for consequential actions, and full tracing. The next stage uses two to five agents with explicit task boundaries, least-privilege access, and automated evaluations on real traffic. The final stage adds policy as code across the fleet, along with canary rollouts, kill switches, inventory, and observability.