Managing scale in the age of agentic AI: 14 problems to solve

In the coming years, enterprises will deploy 50 to 80 times more AI agents than human users. This isn’t incremental change – it’s an identity tsunami.

Today’s agents aren’t just bots or microservices. They are autonomous, goal-directed AI entities acting on behalf of your organization. They orchestrate tasks, call APIs, manage data pipelines, and collaborate with other agents at machine speed.

This brings profound new security, compliance, and operational risks that must be addressed.

Problem 1: The speed mismatch

• Agents operate at microsecond lifecycles, spinning up and down in milliseconds.
• Human-centric IAM and IGA operate at day-scale speeds, leaving a critical gap.
• Why this matters: Without systems designed for agent speed, approvals, policies, and audits become obsolete by the time they are enforced.

Problem 2: Lack of visibility and observability

• Most organizations have no real-time visibility into:
  • Agent population
  • Lifecycle state
  • Scope of privileges
• Why this matters: You can’t secure what you can’t see.

Problem 3: Privilege risks

• Many agents operate with overly broad OAuth scopes or static credentials.
• Creates silent privilege escalation paths, where agents gain unintended access to sensitive systems or data.
• Why this matters: Standing privileges and over-scoped tokens are a leading cause of breaches in machine and human identity systems alike.

Problem 4: Fragmentation and silos

• Agents are spread across:
  • Google Vertex
  • Azure Agent Foundry
  • LangChain orchestrators
  • MCP (Model Context Protocol) frameworks
• Each enforces its own local policies, but no centralized governance or risk scoring exists.
• Why this matters: Fragmentation leads to inconsistent controls, unmanaged shadow agents, and security blind spots.

Problem 5: Unregistered shadow agents

• Agents running outside approved workflows or registration:
  • Increase attack surface
  • Create operational chaos
• Why this matters: Shadow agents bypass policy enforcement and monitoring, undermining Zero Trust architectures.

Problem 6: Zombie credentials

• Tokens or credentials issued to agents that are no longer active remain in circulation.
• Creates orphaned privileged identities with no ownership or accountability.
• Why this matters: Zombie credentials are a persistent privilege risk vector.

Problem 7: Privilege escalation

• Over-scoped tokens grant agents access far beyond their intended purpose.
• Why this matters: Violates least privilege principles and increases blast radius if an agent is compromised.

Problem 8: Compliance failures

• Lack of audit trails for agent actions violates:
  • SOX
  • HIPAA
  • Emerging AI governance regulations
• Why this matters: Without evidence of controls and traceability, compliance cannot be demonstrated.

Problem 9: Data leakage

• Agents can exfiltrate sensitive data by interacting with APIs and tools without policy-aware controls.
• Why this matters: Undetected data leakage undermines privacy obligations and security posture.

Problem 10: Operational fragility

• No consolidated view of agent population prevents:
  • Effective troubleshooting
  • Performance optimization
  • Incident response
• Why this matters: Undetected misbehaving agents can cause outages, performance degradation, or cascading failures in automated workflows.

Problem 11: Incomplete agent discovery

To manage agents effectively, discovery must cover three categories:

1. Platform-Resident Agents
   • Agents running in frameworks like LangChain, Azure Agent Foundry, or Google Vertex AI.
   • Often invisible without platform-integrated discovery.
2. Ad Hoc Inbound Agents
   • Agents arriving dynamically via MCP-based APIs or delegated execution frameworks.
   • Appear transiently, leaving no trace without real-time registration and monitoring.
3. Runtime Agent Observability Logs
   • Agents generate ephemeral logs containing:
     • Execution context
     • Delegated tasks
     • API calls
   • Without capturing and integrating these logs, critical audit trails, behavioral data, and risk scoring inputs are lost.

Problem 12: Absence of an agent fabric

• Identity Fabric unified human and machine identities across clouds and vendors.
• Agentic AI demands a parallel construct: an Agent Fabric.

What’s missing today:

• A registry to discover and register all agents
  • Platform-resident, inbound ad hoc, and ephemeral runtime agents.
• Metadata tracking:
  • Lifecycle, owner, provenance, TTL, execution context, behavior.
• Runtime observability log ingestion
  • Normalizing logs from orchestrators, MCP endpoints, and runtimes for full visibility and audit.
• Real-time risk scoring
  • Privilege levels, behavioral anomalies, policy compliance, execution patterns.
• Integration with policy engines
  • Enforcing dynamic, context-aware authorization.
• Centralized observability
  • For every agent action, delegated task, and API call.

Problem 13: Lack of Agent risk scores

• Agents lack real-time risk scoring, unlike human identities.
• Why this matters:
  • Without risk scores:
    • Privilege sprawl goes undetected.
    • No step-up controls for high-risk actions (e.g. payments, data exports).
    • Security teams can’t prioritize based on risk.

Problem 14: Treating agents equally as human users

• Treating all agents equally is unsustainable as the number of agents dwarfs human users.
• Why this matters: Without differentiated controls and risk-based policies, organizations will lose control as agent scale explodes.

The question isn’t whether you have an agent problem. It’s whether you even know how many agents are running in your environment right now – and what risks they pose.

The age of agentic AI is here. These problems must be solved to build trust, enforce security, and govern AI at scale.

Join Strata’s Agentic Identity preview.