Last week I went to the DevOpsDays BCN meetup, hosted by Allianz Technology Spain at their Barcelona office. Two talks stood out, so here’s a quick recap for anyone who missed it.
Nobody’s on fire, everybody’s tired
David Castañeda (Senior DevOps Engineer at Xcelirate) and Eduard Epure (Security Engineer at Mews) shared two very different takes on the same question: how much should you trust code you didn’t write?
David runs user-facing infrastructure, where one bad change is visible to customers straight away — so he reads every line an AI writes and pushes back often. Eduard builds an internal platform used by a couple hundred engineers, and ships AI-generated changes once a pipeline of syntax checks, security checks and an AI reviewer goes green, spending his attention on running specs in parallel instead.
Two senior engineers, two companies, two opposite answers. What I took away: the right amount of trust in AI-generated code depends less on the model and more on the blast radius of what you’re shipping.
Deploying Agentic AI with GitOps, without reinventing the wheel
Pere Alcoberro Turu (Head of Platform Engineering at Allianz Technology) showed how to treat AI agents as just another production workload instead of inventing a brand-new delivery model for them.
Using AWS Bedrock AgentCore, Crossplane and ArgoCD, he reused existing Crossplane compositions built for ECS-based applications and adapted them to AgentCore’s requirements — modeling the agent runtime, networking, IAM and supporting AWS resources behind a standardized API, with Git as the source of truth and ArgoCD doing the reconciliation.
The core idea resonated with me: Agentic AI doesn’t need a special deployment process. Treat agents as applications, expose their infrastructure through reusable platform APIs, and a pull request becomes the interface to your deployment platform — same as everything else.
Takeaways
- AI-assisted code review should scale with the impact of what you’re deploying, not be a blanket policy.
- If you already have solid GitOps and Platform Engineering patterns, agentic AI is usually a new workload type, not a new platform.
Thanks to the organizers and to Allianz Technology for hosting. Looking forward to the next edition.