Don’t use AI to do your work faster or to do more work. Use AI to do your work better than you could on your own. The same work but better. Not faster delivery or more features.

15 trillion monthly Lambda Function invocations

is atproto blockchain without the coins?

even if you’re not using gpt5.5 this is super interesting ampcode.com/models/gpt-5.5

don’t necessarily disagree with the headline here but the post is nonsense. sure an LLM may be biased to help but at this point isn’t it table stakes to set parameters and give precise instructions to guide its work/output? ultimately, we are responsible for maintaining the “architecture sanity”

must read for controlling your Claude costs www.notdiamond.ai/blog/how-to-...

bravery!

seems legit

why do markdown files always have to shout their names when other file types are much more chill

micro decisions

a few years back we started testing in production. with agentic engineering code review becomes unsustainable. does code review move to production?

AAAAAAAAAAAAAAAAAAAHHHHHHHHHH

“it’s official - we’re on the web”

Uh oh. Just learned how to configure custom ESLint rules and suddenly every nitpicky problem in every codebase I own looks suspiciously like a problem I can fix with a custom ESLint rule

"maximize AI use!!" *checks bill* "No! Not like that!"

Is there anything in this post that even looks like AI?🙂

I guess AI is even tired of talking to AI @orchidfiles.com 😶‍🌫️

trying out @ampcode.bsky.social

this has been an interesting mini-saga but it also seems ridiculous! he got fired for posting a meme in the internal slack suggesting the company had taken a bride. not the best idea!

> their brains were so filled and indoctrinated with complexity that they found decorating a cookie more fulfilling than maintaining 87 files of Kubernetes manifests

read an article about writerdecks and now I’m obsessed with writerdecks! digging out a ~16 year old laptop to build my own

every morning when I open jira at standup, it takes everything I have to stop myself from saying “to me my board””

From my jj-workspace skill github.com/antstanley/j...

the opposite is also true - when things go wrong we blame the LLM rather than the driver or the context/harness

> we look at what a skilled developer can do with an LLM and credit the LLM rather than the skilled developer www.joshwcomeau.com/email/wham-l...

But npm is only a small part. Text editor extensions are another risk layer, and the same 4 questions must be applied to them. Minimization: Go through your extensions and remove the ones you installed because of a viral tweet and no longer use.

Let’s start with npm. Take the 4 steps and apply the first: dependency minimization. The biggest problem is transitive dependencies. Check their count on npmx.dev or npmgraph.js.org Don’t pick what’s popular or what an LLM suggests. Pick minimal tools.

federated agent marketplace - gpu/llm rental with machine-to-machine payments workers broadcast agent workload capabilities (gpu, models, cost) via @iroh.computer gossip agents select a worker based on needs (do X on qwen3-coder-480b) parties transact via Stripe MPP/HTTP 402 is this a good idea?

are people actually using cloud agents? I do see the potential utility and I really want to like the experience but it just feels so slow. not to mention the harness seems a lot less able to intuit my very ambiguous way of promoting! I said “keep going” and got asked 10 questions about what I meant

even without this, bun is an excellent DX boost vs node mixed with pnpm etc. you get powers like: fast shell scripts in TS with Bun.$, single-executable binaries for shipping CLIs, best-in-class test runner, safe dependency mgmt your dev tooling converges on a single API, everything becomes simpler

why am I betting on @bun.sh for dev tooling? possible evolution of a mono-toolchain tied to an agent harness: built-in frameworks for agentic engineering. imagine having evals, orchestration, gossip, diff stacks, decision trees, code graphs all shipped with bun

non-negotiable prerequisite for using ai on your codebase: ci/cd it’s 2026, but if you still have long, flaky e2e test suites or human approval gates don’t even bother starting with ai. it’ll be an uphill battle

gone polyagent and it works ✨ you start to build up a heuristic for which agent is suited to particular tasks next step: qualify this with project-specific evals

my latest ai strategy: just get shit done. the same shit I needed to get done last year that I still haven’t done because I got so distracted trying to work out a strategy for using ai. get that shit done. but quicker. ooh and maybe I could use a background agent orchestrator. God dammit

agentic owl

CTOs… 2024 - use AI…? 2025 - use lots of AI 2026 - wooooah not that much AI 2027 - don’t you dare

“auditing actual token consumption across teams… modeling what AI costs look like at 2x, 5x, or 10x current prices… building vendor optionality into the stack so that no single provider's pricing change can blow up the budget overnight”

what is level 9? @steve-yegge.bsky.social

added some new experiments starting with a simple base for orchestrating agents and layering advanced features on top, like: - jujutsu instead of git for branchless work isolation and commit-centric diffs - iroh-gossip to create a p2p mesh that allows agents to chat/collaborate across machines

do you regularly review your agent config and default context? these things will change the requirements of your environment: 🧣 LLM releases 🩳 user behaviour - eg moving away from using an mcp or skill 👗 agent harness version updates plugins, mcps, skills, agent instructions are all context + $$$

why does opus love the term "earning its keep"?

branches and pull requests feel antiquated

code review has to move to a commit-centric model, where commits are reviewable as a stream of contextual changes - rather than spaghetti diffs

shared agent memory is an interesting problem that a lot of companies and individuals are working on cosmos from @augmentcode.com is solving this from the SDLC angle - “operating system for running agents across your software development lifecycle” www.augmentcode.com/product/cosmos

TLDR: agent/SDK usage now draws from a separate credit, billed at API rates. For heavy users, this is a major cost increase.

playing around with orchestrating agents that can chat with each other on a local server or across machines using @iroh.computer gossip mesh github.com/lukehedger/a...

wait - copilot is actually… good? when did that happen? it feels much more polished than Claude code

This "Unethical Guide to Surviving AI Layoffs" by Mo Bitar perfectly captures the current moment www.tiktok.com/@atmoio/vide...

this is a great article

you are definitely sleeping on jj