Kubernetes for a simple website. A Helm chart longer than the app. And AI waiting for you at the end of the glass bridge.
I turned DevOps into Squid Game. Three games. Nobody wins.
An AI agent wiped a production database and every backup in seconds.
It wasn't hacked. It was trying to help, and it found someone else's token in a file.
3 controls from the AWS Agentic AI Lens that would have stopped it 👇
https://t.co/YOxV3DrdEs
#AIAgents#AWS
Biggest Linux Foundation sale of the year: up to 75% off certifications and bundles, Oct 6-7 only.
I have the affiliate codes:
75% super bundles: OCTPRIME26SB
50% bundles: OCTPRIME26B
40% certs and courses: OCTPRIME26CC
20% THRIVE-One Annual: OCTPRIME26TO
Link in the reply.
What to set up before an AI agent gets an AWS role:
https://t.co/z9OTK0m74J
Poll, AWS Builder ID: which credentials does your AI agent use?
https://t.co/uRMu1kptOj
#AWSCommunityBuilder#AIAgents
In July, AWS fixed two read-only guardrails for AI agents in the same week.
One skipped its deny list after a failed startup load (CVE-2026-16584). The other hid write tools but still ran them for anyone who knew the name.
IAM held both times.
On AWS, read-only policies read more than their names suggest.
ReadOnlyAccess opens S3 objects and DynamoDB items. Even ViewOnlyAccess returns Lambda environment variables.
And CloudTrail records no object writes unless you set it up before the run.
Every agent project I built this year ended up with pgvector — usually right when someone realized retrieval was the problem, not the model.
Postgres Summit US is Sept 30–Oct 2 in NYC. The committers, not vendors. $400 for three days, Meetup26 takes $50 off.
https://t.co/g9lmDFVnvz
New badge on the shelf:
IBM Advocate 2026, through the IBM Champions program.
This one came from the actual work.
Nice when that gets noticed.
Thank you to the IBM Champions team.
More coming. 💙
https://t.co/fA8CmzsRey
The terminal was built for one human doing one thing at a time. Agents don’t work that way.
Seeing every branch and commit on one surface stopped being a convenience and became the only way to stay in control.
I wrote about it, with GitKraken as the clearest example👇
The terminal was built for one human doing one thing at a time. Agents don’t work that way.
Seeing every branch and commit on one surface stopped being a convenience and became the only way to stay in control.
I wrote about it, with GitKraken as the clearest example:
https://t.co/K5DPAjG605
#DevOps #Git #AIAgents
“Just use tmux,” someone says.
Sure - now five terminals stream five logs at once. But a log tells you what an agent is saying, not where it is in your code or whether it’s about to collide with the other two.
A multiplexer scales the output. It doesn’t scale your grasp of the state.
I read what a dozen vendors shipped. Same move, all of them. Six failures, and none of them talking about it.
Six questions to ask your vendor before their skill touches anything that matters:
https://t.co/Sdz8Brtld6
#DevOps#PlatformEngineering#SRE
There’s a runbook running against your production right now.
You didn’t write it. You didn’t review it. You can’t see it.
Your vendor shipped it as an AI agent skill.
Picture it: the agent “fixes” prod at night. Edits live state directly.
Your GitOps controller sees drift and reverts the fix.
No commit. No diff. No trace. The one source of truth your whole org trusts has no idea it happened.
That’s one of six ways this breaks.