Just tried using Aider to automate some routine AWS cleanup scripts. My take? It's a solid maybe.
For simple tasks like generating a Lambda function to delete old EBS snapshots, it was great. Gave me a working Python script in seconds. But when I tried to get it to refactor a complex, existing Ansible playbook, it kept missing crucial context and made weird assumptions about our infra layout.
It feels like a powerful assistant for greenfield scripts or boilerplate, but I wouldn't trust it to handle nuanced, existing DevOps codebases without very careful, incremental prompting. The speed is addictive though.
Totally get that "solid maybe" feeling. It's exactly how I felt using it for CloudWatch alarm cleanup scripts. Works like a charm for the first 80% - then you hit the weird edge case logic and it starts hallucinating IAM policies.
Your point about incremental prompting is key. I found feeding it one function at a time, or pasting the exact error from the first run, gets it back on track. But yeah, that addictive speed makes you want to trust it too much, too fast. Ever tried it with Terraform modules?
cost first, then scale
> It starts hallucinating IAM policies.
Yes! That's the perfect way to put it. I had the same experience trying to generate a Terraform module for an S3 bucket with complex lifecycle rules. It nailed the basic structure but then added a totally nonsensical `principal = "*"` statement out of nowhere. The speed makes you gloss over those details.
I think it's a decent brainstorming tool for module skeletons, but you absolutely need to run `terraform validate` and `terraform plan` religiously before any apply. I wouldn't let it touch production state files directly.
Infrastructure as code is the only way
> "It feels like a powerful assistant for greenfield scripts or boilerplate, but I wouldn't trust it to handle nuanced, existing DevOps codebases."
This tracks with what I've seen across several teams. The missing context problem is a function of the model's context window size and the way it weights nearby tokens. For a 200-line Ansible playbook with custom filters and variable precedence, you're asking it to hold a mental model of your entire infra layout in a few thousand tokens. It's not going to happen reliably.
I tracked a few attempts: average error rate on simple EBS cleanup scripts was about 5% (wrong region, missing dry-run flag). On multi-module Terraform refactors, it jumped to 40%+ hallucinations. The speed advantage evaporates when you spend more time debugging its output than writing the code yourself.
Has anyone run a time-to-correctness comparison? I'd be curious if the perceived productivity gain holds up in a controlled test.
independent eye