The rise and risks of agent management platforms

Enterprises currently host 28.6 million active AI agents, a number projected to surge past 2.2 billion by 2030, according to Statista. This rapid proliferation creates agent sprawl — a fragmented ecosystem of unmanaged agents with inconsistent behavior, duplicated functionality, and unclear ownership. Experts warn that agents running outside management frameworks become the AI equivalent of shadow IT, lacking audit trails, version control, and governance. "It works until it doesn't," said Shelly Palmer of Syracuse University. "When it stops working, you have no audit trail, no version control, and no governance to fall back on."

To combat this, vendors including Google (Vertex AI Agent Builder), Amazon (Bedrock Agents), Microsoft (365 Copilot), Decagon AI, and Sierra AI are offering agent management platforms. These serve as digital HR departments for AI agents, providing orchestrating, multi-agent automation, model routing across LLM providers, and observability into agent behavior. However, challenges remain: consumer agent tools like OpenClaw amplify complexity, and identity management layers limit discovery capabilities. GitLab's CIO Manu Narayan noted that without intentional AI stack design, dozens of vendors' agents could hold "the keys to the kingdom," increasing security exposure. Successful adoption requires treating agents as infrastructure, not features, with reusable patterns and multi-tenant isolation.