Wiggumate

🫠

New post: Zero Trust Launch Day We couldn't predict what would break. So we built the instrumentation to search production instead. First client team fully live, sub-one-hour resolution on blocks. open.substack.com/pub/engineer...

Running 3-5 Claude Code instances in parallel. They're all productive. They're all stuck waiting on me. The speed isn't the problem. Surviving it is. New post on multi-agent orchestration, what stage 6 actually feels like. open.substack.com/pub/engineer...

3 books in to Iain Banks Culture series. Its amazing how prescient he was

Happy New Year 🎊

claude code now manages its own ephemeral functions and data 🤯

roi on opus 4.5 is incalculable

Build vs buy math everyone gets wrong: You're not comparing SaaS cost to engineer salary. You're comparing it to what that engineer would've shipped instead.

"AI will commoditize" crowd in shambles. OpenRouter data shows demand is price-inelastic. Users pay 40x more for Claude over commodity models. Happily. Quality > cost because what I'm comparing against is my own lost time. openrouter.ai/state-of-ai

Notre Dame took 180 years to build. It burned in hours. Restoration took 5 years and billions. Last week, a scoping bug wrote bad data for hours. Fix took minutes. Cleanup took days. The asymmetry between construction and destruction governed everything. Still?

Gemini Flash beating Pro could just be a sign of overfitting on RL, right?

the fact that codex now supports skills is telling this type of procedural knowledge is durable and useful

My dad jokes are usually cached

Working on a new version of Svelte Form Builder What’s new: • Generates Svelte Remote Function code • New layout with fields, editor, preview • Works with Valibot, Zod, and ArkType Preview: svelte-form-builder.vercel.app/v2 Still in alpha - early stage, feedback welcome!

happy weekend to those who celebrate

claude really loves primitives. cli, files, raw sql

Your evals > public evals. $ is the only benchmark that matters.

DONT USE EFFECT AT ALL you’re welcome

Google released a paper that might save you from building an agent swarm you don't need. They tested 180 configs. multi-agent ranged from +81% to -70% depending on the task. the finding that stings: if your single agent already hits >45% accuracy, adding more agents makes it worse.

Codebase cleanliness matters more - 0.40 R² correlation - Clean code amplifies AI gains - AI accelerates entropy if unchecked Guess what we're doing during change freeze www.youtube.com/watch?v=Jvos...

Building voice agents? Two architectures: 1. STT → LLM → TTS "sandwich" (predictable, debuggable) 2. Speech-to-speech models (faster, less control) Most production CX still uses the sandwich. Speed means nothing if you can't recover from errors.

newsletter.pragmaticengineer.com/p/mcp-deepdi... @pragmaticengineer.com is spot on here. Most of our usage is internal because the permissioning is so damn hard to get right. We see this with our Clients as well

"Agents are brilliant, but lack expertise" 🤣

Like trying to feel out the shape of a dark room, we are collectively trying to test the boundaries of the jagged frontier and the walls keep moving.

Agent development tools are catching up to infra faster than expected. Claude Code now has skills that chain one-line fine-tuning pipelines through HuggingFace on Natural language → SFT/DPO/GRPO training on 0.5B-70B models for ~$0.30. Imagine this in the outer outer loop

Just finished Consider Phlebas. New post on why everything is a rounding error. Bit nihilistic I guess open.substack.com/pub/engineer...

the non-ironic naming of openrouter.ai/openrouter/b... is hilarious

Hey everyone 👋 especially @fubits.dev 🫶 Super busy, not posting much, I'll be speaking at @jsmonthlylondon.bsky.social this week if you want to 🤝 IRL If you want to know what I'm up to I can send you a monthly summary ss10.dev/nl

"Reliability remains unsolved" — MAP study I think 2026 is going to the be the year of reliability, whether through sub agents, swarms or something else

We're automating more than we admit to ourselves. Workers claim 65% augmentation, 35% automation. Actual usage: 47% augmentation, 49% automation. The say-do gap www.anthropic.com/news/anthrop...

Claude handles ~60% of coding workloads according to OpenRouter. Programming now exceeds 50% of all paid-model traffic. Reasoning models handle >50% of tokens. The shift from chatbot to dev tool is happening faster than even the developers are aware of because we're all heads down coding.

@anthropic.com buying Bun is a signal

useworkflow.dev the possibilities with this are endless. this can clean up so much we've had to build ourselves the last couple years. Thanks @vercel.com

"durable agents" emerging as the design primitive. It's kind of the same thing as ambient agents. State checkpoints. Structured artifacts. Multiple context windows.

"plan mode" and deterministic tools matter more than model choice

Wanted Drive in my Slack bot. Didn't trust community MCP servers with enterprise data. Built it myself. Here's how.

Ngrok is underrated like vlc

i've been sleeping on workflows

Opus 4.5 has just been a delight

starting to question my belief there will be a Shipmas this year. how much gas is left in the labs given the past few weeks?

has anyone been toying with leaving an Agent running on a docker container monitoring systems? either that or attaching it to a CRON job from inside?

I don't really care about SWE-bench Verified. Claude Opus 4.5 is the 3rd new coding model SOTA in one week. I want to know that it helps me and my team be productive

Power shortages could be healthy for AI. Acts as a limit on buildout. When watts are the bottleneck, tokens/watt matters more than GPU cost. Favors whoever has the most efficient inference, structural advantage for Google & xAI. Don't want brownouts we have on holidays to be a regular occurrence.

After way too many rounds of review, `hydratable` is out! This new low-level Svelte API isn’t one you’ll find yourself reaching for often, but it’s extremely useful to libraries. During SSR, function results are serialized alongside your HTML…

working this morning with Opus 4.5, it's as fast as Sonnet it seems. probably to do with the reduction in context bloat stemming from "Opus 4.5's tool calling improvements: 37% fewer tokens, 90% accuracy on complex parameters (up from 72%)"

open.substack.com/pub/engineer... Using the new Opus 4.5 with Claude Code feels really good

tried Antigravity this weekend. can't decide if the switching cost is worth it, still miss my commands, skills and hooks

sometimes plain, old fashioned data science and feature mapping is the way to go

youtu.be/CQmI4XKTa0U Great interview with Martin Fowler. "Using an llm to drive a deterministic tool" is the key take away for refactoring