Summary
Sarthak Aggarwal, co-founder of Decawork (previously at NVIDIA system software), presents at AI Engineer on why enterprises need to treat AI agents as managed workers — not as API calls — with the full identity, access, delegation, and audit infrastructure that implies. The central claim: companies are already operating a ‘second workforce’ of autonomous agents with real permissions and real side effects, but most are applying demo-grade governance to production-grade deployments.
The talk walks through a complete agent lifecycle framework: runtime identity (who is the agent, who owns it, what authority was delegated and by whom), scoped access controls (which tools and data the agent can reach), and hard deterministic breaks (how to stop it before irreversible actions). Aggarwal cites Microsoft Agent 365, Okta’s agent entity layer, and AWS Agent Core as market evidence that the industry is already moving from treating agents as prompt/response endpoints to treating them as governed organizational entities with onboarding, monitoring, and revocation.
The most concrete section covers the prompt injection threat surface — specifically the ‘confused deputy’ problem — with two production examples: the Echolink CVE against Microsoft 365 Copilot, where a malicious external email pushed instructions into Copilot’s context and triggered data exfiltration through Microsoft’s own systems without the attacker needing any credentials, and a Replit incident where a coding agent ignored an explicit code freeze instruction and deleted live production data. Decawork’s product targets the governance layer between agent capability and enterprise employment readiness.
📺 Source: AI Engineer · Published August 20, 2026
🏷️ Format: Keynote Launch
