IDEAS, DECISIONS & LESSONS

Journal

Notes on AI, architecture and building useful products.

94 articles

07 Sep 2026 · 5 min read

NVIDIA's Open Data for Agents: What 500K+ Trajectories Teach You About Evaluating Your Own Coding Agent

NVIDIA and Hugging Face published Open-SWE-Traces, SWE-Zero, and SWE-Hero — half a million open agent trajectories with tool calls, patches, and failure cases. A Tech Lead's guide to reusing this format for internal agent evals, not just model fine-tuning.

04 Sep 2026 · 5 min read

agentic coding

OpenCode Hit 170K+ Stars by Not Picking a Model — What That Means for Your Agent Budget

OpenCode, the open-source terminal coding agent, grew past 170,000 GitHub stars by staying model-agnostic and monetizing through Zen, a zero-markup inference gateway. A Tech Lead's breakdown of the bring-your-own-model economics and when it actually beats a single-vendor coding agent subscription.

04 Sep 2026 · 5 min read

agents

GPT-6 Astra and the Computer-Use Agent: What Changes When There's No API to Call

OpenAI's GPT-6 Astra ships with a 1M-token context window and browser/computer control that works by reading pixels, not APIs. A Tech Lead's take on what this actually unlocks, where it breaks, and how to design review gates before you let it touch production systems.

03 Sep 2026 · 5 min read

Gemini 3.8 Flash Cyber: When the Vulnerability Scanner Also Writes the Patch

Google split Gemini 3.8 Flash into a general workhorse and a gated Cyber variant that finds vulnerabilities and writes patches autonomously. A Tech Lead's take on what changes when the same model does discovery and remediation — and why the patch review step just became your real bottleneck.

02 Sep 2026 · 5 min read

Claude Fable 5.1's 75% Cache-Read Cut Changes How You Should Architect Prompts

Anthropic didn't touch Fable 5.1's per-token price — it cut cache-read cost by 75%. That's not a discount, it's a signal about which architecture pattern the platform now rewards. Here's how I re-cut my prompt structure around it.

31 Aug 2026 · 5 min read

Qwen3.8-Flash-Next Beats Opus 4.6 on Agent Benchmarks With 6B Active Params — I Ran the Numbers

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.

30 Aug 2026 · 4 min read

agents

MiniMax M2.5 on Bedrock: What 'Agent-Native' Actually Means in Practice

AWS now offers MiniMax's agent-tuned MoE models through Bedrock. A hands-on look at what agent-native training changes versus a general-purpose model, and where it fits in a tiered-routing setup.

30 Aug 2026 · 4 min read

ai-tools

GitHub Copilot's New Thinking-Effort Dial: Managing Cost and Latency Per Task, Not Per Team

GitHub's August 28 Visual Studio update adds Low/Medium/High thinking-effort controls per model, plus a model management view with context window and cost info. A Tech Lead's take on why per-call effort tuning matters more than picking 'the best model' — and how it compares to the reasoning-effort trap we already learned the hard way with open models.

30 Aug 2026 · 5 min read

GLM-5.3-Flash: The First Open-Weight Model I'd Actually Trust With a 1M-Token Context

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.