@ibuildthecloud The new code review process:
Agent: Here’s my plan.
Me: Sounds reasonable.
Agent: changes 47 files
Me: I should probably start reading the plans.
There’s something appealingly Unix-like about this.
Instead of building an increasingly sophisticated “memory system” that guesses what matters:
append everything → search when needed.
Sometimes the better abstraction isn’t smarter memory.
It’s cheaper access to the original state.
Most AI workflows are held together by prompts, retries, and hope.
But unreliable components, cascading failures, and poor observability aren't new problems.
I explored applying distributed systems principles and evaluation-driven development to AI workflows.
https://t.co/xVM6GrQh7e
AI can now build five microservices, configure Kafka, and deploy everything to Kubernetes in minutes.
But what if you only needed a Postgres table?
We've made bad architectural decisions cheaper to execute. Not cheaper to maintain.
My latest post:
https://t.co/pDYZZvY1Tn
A strange thing about working in tech right now:
The half-life of being “experienced” is getting shorter.
You can have 10 years of experience and still wake up to a world where the tools, workflows, and expectations changed underneath you.
Experience still matters.
But only if you’re willing to become a beginner again.
Happy Choti Diwali 🪔
May your lights be bright, your sweets be unlimited, and your production alerts remain suspiciously quiet.
Some things are better left unlit. Especially PagerDuty.
What if your AI agent doesn't need a better model, just a better workflow?
A new paper, Strategist, explores agents that generate, test, and optimize their own reasoning strategies.
It reports comparable accuracy at 39% lower inference cost.
Maybe agent architecture shouldn't be designed once. It should be continuously evaluated and optimized.
https://t.co/hYAjbWx7e2
The next big improvement in AI might not come from a bigger model.
It might come from making attention cheaper.
A new paper on Latent Sparse Attention demonstrates efficient attention over contexts reaching 10 million tokens.
The interesting part? Instead of searching through every token, it clusters similar information and searches the relevant groups.
Think indexing instead of scanning.
For AI coding agents, this could mean keeping entire repositories, debugging histories, and architectural discussions accessible without constantly compressing context.
But there's a catch.
Making information easier to access doesn't make it more reliable.
We may be solving the context scaling problem faster than the context correctness problem.
Paper: https://t.co/Rlq7moDht7
AI won’t eliminate software engineering teams.
It will eliminate the justification for bloated ones.
A team of 5 strong engineers with AI leverage may soon outperform teams of 20 built around coordination, handoffs, and specialization.
The next scalability problem might be the org chart.
Somewhere along the way, we started treating every free hour as an opportunity to be productive.
Read more. Learn more. Build something. Earn something.
Maybe a Sunday spent doing absolutely nothing is a perfectly good Sunday.
I’ve started documenting my Kubernetes Cluster API learning journey.
First up: understanding the core idea behind CAPI, using Kubernetes-style declarative APIs and controllers to manage Kubernetes clusters themselves.
Notes from what I’m learning here:
https://t.co/msheahKfSA
Pods are great execution units.
But for AI, HPC, and tightly coupled distributed workloads, they’re often the wrong scheduling abstraction.
Kubernetes is moving toward workload-aware scheduling with PodGroups, gang scheduling, topology-aware placement, and workload-level preemption.
I wrote about why that shift matters:
https://t.co/b3qH85Gz5w