Conversations on controlled autonomy

The autonomy envelope: drawing deterministic boundaries for AI agents

Venkat Siva (Sivasubramanian), CEO of CompFly AI, with Lockhart and Gary Chan

When a human pulls a CRM record, security teams look for human context to proxy intent — unusual hours, sudden bulk exports, or HR signals. But when an AI support agent runs that exact same sequence, those human signals disappear, and we end up treating autonomous code execution like standard, predictable API traffic. The primary risk in agentic AI is not just prompt injection: it is an agent chaining together legitimate permissions to create an unintended outcome at machine speed. You cannot catch this with a basic filter — you need a clearly defined autonomy envelope to establish deterministic boundaries. Reading an audit log after the fact is not enough; you need a dedicated control plane.

• Identity: giving every agent a verifiable identity tied to its function
• Pre-flight: running structural evaluations before pushing to production
• Runtime: enforcing deterministic guardrails without a massive latency tax

Governing autonomous agents at the execution boundary

Venkat Siva (Sivasubramanian), co-founder & CEO of CompFly AI, with Professor Kevin Werbach

Your AI model can be perfectly safe, yet your agent can still cause real harm. Venkat Siva joins Professor Kevin Werbach to discuss why governing autonomous agents demands a new control layer rather than a feature bolted onto existing security tools: why governance has to live at the "execution boundary" where decisions become actions; the limits of least privilege and "measured autonomy" as an alternative; giving agents cryptographic identity, provenance, and a tamper-evident trust ledger; "shadow agents" and how enterprises can even discover what is running; setting an "autonomy budget" so agents aren't an expensive suggestion box; and what changes when agents move into physical systems like cars and robots.