Aloha!
There’s so much bullshit about AI circulating that it personally makes me want to vomit at least once a day. But under all this there are nuggets. As I’m spending time on this and information sinks in I can feel how this will change platform engineering.
I believe AI will affect us from two angles: first, it will influence the tools at the disposal of developers and second it will influence the type of applications we’re building.
I will not bore you with all the different directions this could take but here are a few observations that could be interesting:
- Human out of the loop executions in the enterprise should not exist until you cannot guarantee close to 100% accuracy. Tolerance for failure in a production environment should be zero. This is similar to the self-driving car problem. If you ask a system to write your Terraform to create S3 and it’s right in 98% of cases, you’ll still need to read every file. You win relatively little.
- While the above analysis is correct it doesn’t change the fact that people will still do it and they are already doing it. And yes, in the regulated enterprise.
- Human in the loop, specifically looking at the tab functionality in cursor for instance is the much better option. The optimization is essentially, to not predict the next 100% but the next 20 characters and let the human in the loop confirm the change. This takes longer but if you include the time spent on reviewing machine-created code it’s still faster.
- Whatever strategy you opt for you need to catch faulty stuff more accurately before it hits the infrastructure. Policy checking gates become increasingly important in this world. I would expect Snyk or OPA to become even more important than they are today.
- The agentic workflow case beating all else is not one I totally buy. Translating human input into calling a backend API is not something that I would call agentic, most approaches I see so far are still just BPMs with LLMs in the nodes. Certainly helpful for some tasks but often more smoke than mirrors.
- In order to use AI, be it in agentic workflows or to create infra code, having a strong backend API is absolutely necessary. And that’s a.) to receive the request translated by the agent and b.) to match the request with the default for how infra should be configured in a given context to fit the policy of the organization.
Finally it will be interesting to see how we’ll deal with all those creative landing pages that marketing teams bash out. Which are 100% auto-generated and will still need to be hosted. That will be fun, enjoy yourselves!
I’m very excited about this topic and will be spending a lot of time on how to debug this new world.
Have something to say about platform engineering? The call for proposals to speak at PlatformCon 2025 is closing in just a few weeks. Submit something!:)
Enjoy yourself and see you soon!
Kaspar
