Ugur Basak
Speaking
Talk ·
AI Agentic Systems Platform Engineering Resilience Security

Resilience by Default

RBLN Europe · Amsterdam

The resilience patterns that kept the software lifecycle standing under load, blast-radius containment, golden paths, and idempotent recovery, applied to AI and agentic workloads, and where the failure modes are genuinely new.


Agentic systems are not a new universe. Almost every failure mode they bring has a twin in ordinary software delivery, and the engineering discipline that contained it already exists. What changes is scale: an agent runs the loop thousands of times an hour, with real credentials, unattended.

This talk takes the three resilience patterns that kept the software development lifecycle standing under load and applies them, directly, to AI and agentic workloads.

Event

RBLN Europe 2026 (the Rebellion security and AI series) runs 2 to 4 September 2026 at Hotel Okura, Amsterdam. This session is on the Main Stage, Ballroom I/II, on Thursday 3 September, 10:30 to 11:00. Full agenda and registration: rbln.com/events/2026/europe.

The core argument

A non-deterministic action is a flaky test. An over-scoped agent token is the credential leak we spent a decade learning to prevent. An agent calling an agent in a loop is a fork bomb. The principles do not change. The blast radius gets bigger, and a small number of failure modes are genuinely new.

What the talk covers

Blast-radius containment. The agent’s blast radius is everything its credentials can reach. Scope every token to the task, make it short-lived, and threat-model as if the agent is already compromised.

Golden paths. Give the agent the same paved road your developers use, where every gate is a deterministic mechanism: identity, policy, a secrets broker, a sandbox, an audit hook. There is no off-road route.

Idempotent recovery. Replay is unsafe when the actor is non-deterministic. Move idempotency into the environment: transactional tool calls, dry-run then diff then apply, checkpoints, and a human gate on the irreversible.

Where it is genuinely new. Prompt injection as control-flow hijack, cascading autonomy that turns your own orchestration into a denial-of-service, and the absence of a clean intended state to reconcile against.

Who this is for

Engineering and platform leaders putting agents into production, and the security and operations teams responsible for keeping them contained, recoverable, and accountable.

Invite me to speak

I speak on Platform Engineering, Developer Experience, Engineering Leadership, and Organisational Transformation.

Invite me to speak