Agent management

"Organizational" structure of agents

As AI agents become pervasive in enterprise environments, a critical question emerges: how will these agents be structured and managed? Contrary to what I see often today, the organization of enterprise agents will (IMO) not follow rigid hierarchical structures or isolated "pyramids" of agents operating independently.

Instead, and similar to the internet, agent-to-agent (A2A) communication will enable broader and more flexible collaboration patterns.

Networks, not hierarchies

While the internet has minimal inherent hierarchy, agent networks may incorporate elements of privilege and authority. Similar to organizational charts in companies, there will be defined relationships determining which agents (or humans) can delegate to other agents (or humans).

These governance structures can be enforced within a company's boundaries but become more complex when collaboration spans across organizational boundaries—mirroring today's challenges when interacting with suppliers or clients.

The most likely outcome is a highly unstructured network of agents and humans. Zooming in within a company, you might see supervisory relationships between agents and humans. However, zooming out to the broader ecosystem reveals a predominantly peer-to-peer style network.

Agent management platforms

A growing field of technology vendors and startups is focusing on enterprise agent management and alignment. These platforms are becoming essential components of AI infrastructure, offering capabilities that facilitate agent discovery, collaboration, and governance.

Trusted agent discovery

One critical function of these platforms is enabling agents to discover each other in a trusted manner:

• "Public" discovery services will likely emerge soon, though these will inherently carry certain risks. Similar to the internet, we will employ PKI (Public Key Infrastructure) for high-trust agents (like commercial banking agents) and "webs of trust" for end-user agents to mitigate these risks.
• Within enterprise boundaries, organizations can maintain private agent gardens, where certificates and other security controls ensure interactions only occur with trusted agents.
• Using public discovery services is like using the public Docker Hub, while private agent directories function more like self-hosted container registries or private artifact stores.

Secure agent collaboration

These platforms also facilitate secure and efficient agent collaboration:

The functionality will resemble API management platforms (like MuleSoft and others), which have been addressing similar challenges in service-to-service communication. It is notable that existing API management vendors are likely to expand their offerings to include agent management capabilities.

Ensuring alignment

Perhaps the most significant challenge for agent management platforms will be ensuring alignment across humans and agents within an organization:

• Just as teams use sprint planning to align human efforts, we will need mechanisms to keep teams of agents and humans working coherently toward shared objectives.
• These platforms will need to provide ways to assign work to agents, visualize current and planned activities, and maintain records for audit purposes.

Looking ahead

As enterprise agent deployment accelerates, the management and governance of these systems will become increasingly critical. The platforms and practices that emerge will need to balance enabling powerful collaboration with maintaining necessary controls and alignment. Hence, I expect Agent Management to become a key component of the AI infrastructure stack.

The agent ecosystem will not be rigidly structured or completely chaotic—rather, it will resemble our existing organizational structures, with defined boundaries within enterprises and more flexible, trust-based relationships across organizational boundaries.