IDEAS, DECISIONS & LESSONS
Notes on AI, architecture and building useful products.
10 articles
03 Oct 2026 · 4 min read
ai
Xiaomi's new open-weight model tops Terminal-Bench 2.1 at 89.9% and jumps from 19.0 to 71.9 on DeepSWE. Before you get excited about MIT licensing, here's what 1.02T total params vs. 42B active params actually costs you in hardware.
Read more20 Sep 2026 · 7 min read
agentic engineering
Two days after writing up Routines, I ran my own morning through Hive: a verified AWS + GitHub brief at 08:30, a feature-planning room, a weekly dependency bump that opens a PR and waits for my click, and budgets at four heights. Here is what ran, what broke, and what I fixed — a dead model alias was killing my rooms without a sound.
Read more18 Sep 2026 · 8 min read
agentic engineering
One day reading n8n, Dify, CrewAI, LangGraph, Paperclip, Lindy, AgentKit and two dozen other agent products to answer one question: what do small teams actually pay for? Not orchestration. They pay for a repeated job that runs correctly, gets checked, and doesn't wreck the bill. Here is the feature I chose to build for Hive, and the list of things I decided not to build.
Read more31 Aug 2026 · 5 min read
ai
Alibaba's Qwen3.8-Flash-Next, an early preview of Qwen4's architecture, scores 62.5 on SWE-bench Pro against Claude Opus 4.6 Max's 53.4 — while activating just 6B of its 180B parameters. Here's what the routing architecture actually buys you, and where the benchmark story gets more complicated.
Read more30 Aug 2026 · 5 min read
ai
Z.ai's MIT-licensed GLM-5.3-Flash pairs a 320B-A18B MoE with a hybrid linear/sparse attention scheme built specifically to make million-token context usable, not just possible. A technical lead's read on IndexPool, the self-hosting math, and where it fits next to closed frontier models.
Read more27 Aug 2026 · 4 min read
agentic engineering
I self-hosted an agent platform on hardware I already own. The models were never the hard part. The hard part was the playbook: the exact combinations of engine, tools, and surface that turn one platform into a coder, a researcher, a QA gate, and a business analyst. Here are the five I actually run.
Read more27 Aug 2026 · 4 min read
agent architecture
Grok isn't a model, it's an agent system: tools trained into the reasoning, real-time X and web search on xAI's infra, a four-agent cross-verifying architecture, long-running agents. I break down its features, architecture, and safety harness — then map each piece to a component I can build self-hosted on a subscription I already pay for.
Read more25 Aug 2026 · 4 min read
agent architecture
I already run a self-hosted autonomous engineer. When I wanted Cursor's cloud agents too, I didn't pick one — I made the executor swappable. Here's the integration architecture that turns any hosted coding agent into a first-class engine behind my own surface, in four small pieces.
Read more21 Aug 2026 · 7 min read
agentic engineering
Devin costs $500/mo. Cursor's background agents live in someone else's cloud with your code. So I built the same capability on my own machines, on a Claude subscription I already pay for. Here is the full architecture -- the loop, the data model, the six-phase build, the hardening -- and why 70% of it was infrastructure I already had.
Read more
