Nek.12

opus/gpt-5.5 can write 8k lines of code in one shot. they can decide which package something belongs in from context alone. you do not need to decompose tasks down to "which line to declare an interface on." that's not helping. that's babysitting.

transferring context between agents and compressing it correctly is a problem the AI agent world still hasn't solved. so why are you fighting the model over it instead of just keeping one session?

models from gpt-5.5 onward don't need this. but ppl are still building on year-old assumptions and paying serious money for the privilege.

context management strategies in harnesses like claude code exist bc older opus lost ~40 IQ points when context filled up. that's where compacts, tickets, step-by-step planning came from. workarounds for model limits.

strict agent workflow = more predictable first draft. my approach = rougher first draft. but i spend 5 min speccing instead of 30+. over a week of work that gap compounds fast.

my agentic workflow: 5-15 min discussing the task → write one text doc → agent works 2-8h and ships it. no swarms, no slash commands, no plan mode. 4/5 failures trace back to a bad spec from me, not the agent picking the wrong package.

agents don't need separation of responsibilities. they need unknown variables removed.

rigid agent workflows strip the model of the flexibility that makes it useful. real tasks always contain something unplanned. something the user left unsaid. a strict system executes to the instructions and produces non-compiling code. then 5 QA cycles. then rollback.

the term is from @steipete and it's exactly right. the tooling nudges you toward this. claude with its agents and workflows practically dares you to add one more layer.

"Claude-pilled" = you've gotten so deep into agent tooling that you believe complexity = quality. it doesn't. frontier models in 2026 are smarter than the systems ppl are trying to pack them into.

@relizarov Builder manages worktrees for me automatically

Trying to manage 3 agents at once. That’s hard. What kind of flow you even use? A new named git worktree for each branch or a set of generic (numbered?) worktrees?

wrote up why swarms, slash commands, and rigid multi-agent workflows are actively hurting you - and what actually works instead: nek12.dev/blog/en/claude…

the more complex your agent setup, the worse your results. fight me.

Claude-pilled: why complex agent workflows are working against you Complex systems of agents, slash commands, and orchestrators aren't progress - they're micromanagement that gets in the way of frontier models doing their best work.... nek12.dev/blog/en/claude…

omg, I was speccing a refactor in Builder and it saw its context usage being high, wrote open questions to docs then triggered compaction on its own. Genuinely insane how smart models utilize their harnesses to the full extent.

@teamcity Tweeting with AI will get you banned on X

1/7 Chapter 1: The Real AI Bottleneck AI has already moved into dev workflows. It writes, debugs, and explains unfamiliar systems. But one part still resists: CI/CD. Not because AI is useless there. Because this is where software stops being an experiment. 👇🧵

@steipete @Cloudflare A roundabout way to do it considering that cloudflare wrangler exists, but I guess you guys were in the flow or needed a high privileged one

Codex complained that one API key didn't had the permissions it needed, so told it to get a new one, it opened @Cloudflare and created a new API key.

I'm again rate limited on GitHub, but codex just opened the browser and clicks around GitHub as workaround.

@getsome_air Queue in 2026.🤦

Introducing Air’s latest feature, the prompt queue. No more waiting just to type the next instruction – queue your next prompts, move on, and Air will deliver them in order when the agent is ready.

@thsottiaux It would be great to know Codex's cache TTLs. If they change often, maybe in response header, or at least a hard coded value if not. I'm building my agent which uses the Codex API and can't do some token optimization magic cause I don't know cache expiration dates.

@jetbrains You even dare to self promote a feature thousands of people begged for for 10 years that takes 30 min to vibecode?

Your coding font might have more features than you’re using. 2026.1 adds support for OpenType stylistic sets in IntelliJ IDEA and other JetBrains IDEs. Configure them in Settings | Editor | Font, and preview glyph changes instantly via a tooltip before applying a set.