IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

Enterprises are deploying a second workforce of autonomous AI agents that act on behalf of users, but the primary challenge is no longer model capability—it is making these agents safe to employ 0:20.

Key Takeaways:
• A working demo proves capability, but not employment readiness; agents with delegated authority require full identity, access, audit, and revocation lifecycles just like human employees 1:45.
• Agents need a runtime identity card defining the actor, subject, delegation context, and exact capabilities, a standard current identity protocols like OAuth do not fully provide 2:31.
• Untrusted text can cause trusted actions, drastically expanding the attack surface—a vulnerability demonstrated by real-world incidents like the EchoLake exploit targeting Microsoft 365 Copilot 6:10.
• Filters and guardrails are insufficient enterprise boundaries; organizations must implement privilege separation where untrusted context is allowed to reason but never exert authority 10:38.
• A robust architecture uses a trusted planner that logs approved actions before seeing any evidence, and a separate executor that processes untrusted data through strict policy gates 12:49.

The future belongs to organizations that apply the oldest enterprise IT playbook—identity, constraints, auditing, and revocation—to this new kind of worker.

Sources:

• 0:20 Introduction of the enterprise second workforce concept
• 1:45 The difference between demo capability and employment readiness
• 2:31 The need for a runtime identity card for agents
• 6:10 How untrusted input creates trusted actions and the EchoLake exploit
• 10:38 Why guardrails fail and privilege separation is required
• 12:49 The trusted intent, planner, and executor architecture pattern