Risk & Resilience in an Agentic world

Agentic systems don't fail like traditional software. When an agent can plan, call tools, take actions and have impact over a wide array of systems. The risk profile shifts from buggy code to unexpected actions in dynamic environments, this increases your blast radius and multiplies harder to predict outcomes. In this session, we'll map the core risk areas unique to agents: runaway tool execution, prompt and data injection, identity and delegated access, hallucinated outputs, fragile multi-step plans, dependency drift and the quiet failures that only show up in production (timeouts, partial completion, silent retries and corrupted states).

From there we'll focus on resilience, the engineering and governance techniques that make agents dependable under real constraints. You'll learn patterns for least privilege tool design, scoped identity and approvals, auditable traces, guardrails that don't stop all interactions and operational tools like red-teaming, incident-ready telemetry.

After implementing production grade, failure-resistant agentic workflows at several enterprise financial institutions, Chris has the deep experience to help you you avoid these pitfalls. You'll leave this session with practical guidance to move faster with agents, without turning your systems, data or brand into the next failure.

A practical map of the top security and failure risks unique to agentic systems (tools, identity, data, and autonomy). Concrete resilience patterns you can apply immediately (least privilege, approvals, guardrails, observability, and safe tool design). A rollout playbook to reduce blast radius in production (evaluation/red-teaming, incident readiness, and governance).