The Blog
Field notes from the build
Deep dives on the AI agents, automations, and developer workflows I build to save engineers hours every week — with the code, the dead ends, and what actually shipped.
Gemini 4 Argon: Fairwind first, API key later
Google shipped Gemini 4 Argon on Sep 30 to Fairwind defenders first. The 1M output limit is real. The API date and the $4/$20 price step are not.
Read →Cloudflare Containers: 648ms starts, and a one-way scheduling change
The durable_object policy picks image and size in code. Median start is 648ms. Snapshots save disk, not process memory, and you cannot flip an existing app onto the policy.
Read →Codex Security Cloud still waits for you to open the pull request
The cloud plugin scans connected GitHub repos, tries to reproduce findings, and will not open a draft PR until you say so.
Read →Cursor Plan Mode: GitHub issue to scoped fix plan in 10 minutes
Use Cursor Plan Mode (Shift+Tab) on a GitHub issue: research, clarify, edit a scoped fix plan, then Build only the first PR slice.
Read →Microsoft just gave vibe-coded enterprise apps a production runtime
Copilot Managed Runtime puts AI-generated apps on a governed Microsoft 365 host — identity, policies, versioning, and admin inventory in public preview.
Read →An Alien Mind and the 3.1 agent-workday: brakes, metrics, shipping calendar
OpenAI's Jakub Pachocki asks for voluntary slowdowns on Sep 6 2026 — same day as 3.1 agent-workday metrics, three days after GPT-6 Astra.
Read →HydraFusion: when Copilot stops choosing a model and starts choosing a workflow
GitHub's HydraFusion research preview reframes routing as orchestration. Read the Sep 4 2026 benches: cost down everywhere, quality clearly up on one.
Read →When local models are ready for day-job agents: Magnitude
When local models are ready for day-job agents: privacy, cost, and offline trade-offs with Magnitude — an inference server that plugs into the agent you already use.
Read →Stop shipping Mermaid slop: Diagram Design for editorial agent diagrams
Diagram Design is an agent skill for editorial HTML+SVG diagrams — brand-matched architecture, flowchart, sequence, and state diagrams without Mermaid slop.
Read →What 'grows with you' means for Hermes Agent
Unpacking Hermes Agent's 'grows with you' claim: memory, skill creation, session search, and user modeling — and where it sits vs harnesses like ECC.
Read →When is a harness OS worth the install tax? Looking at ECC
When a big agent harness pack is worth the context cost: evaluating ECC for skills, instincts, memory, and AgentShield — official channels only.
Read →Humanizer: fixing AI smell without dumbing down the thinking
When the Humanizer agent skill helps polish AI drafts — and when it papers over weak thinking. Install, workflow, and caveats.
Read →Where open-source coding agents fit next to hosted IDEs: OpenCode
Where OpenCode fits next to hosted IDE agents: terminal-first workflow, BYO providers, plan vs build modes, and a practical install path.
Read → SecurityCloudflare OS deep dive: safety, trade-offs, and customizing it on day one
I reviewed open-source Cloudflare OS for call-home behavior, the Gatekeeper security model, and how hard it really is to make it *your* company OS.
Read → SecurityIs diri safe to install? A security review of the multi-agent macOS orchestrator
I audited diri's trust model, install path, remote access, and MCP blast radius — then laid out the field trial for part two.
Read → Prompt EngineeringAI Engineer Coach: the feedback loop prompt engineering needs
Microsoft's local-first coach turns real coding-agent sessions into practical feedback on prompt quality, context, repeatable skills, and healthier workflows.
Read → SecurityIs Alibaba Open Code Review safe for GitHub Actions?
I traced where the code goes, how the Action updates itself, and which GitHub permissions it needs—then built the hardened setup I would actually test.
Read → AI AgentsBuilding my first production AI agent
What I learned taking an agent from a weekend prototype to something I trust to run unattended — tools, guardrails, and the loop that finally made it reliable.
Read → AutomationAutomating my inbox with n8n and Claude
A no-drama workflow that triages email, drafts replies in my voice, and hands me only the decisions that actually need a human. Here's the whole build.
Read →