Blog
Field notes on AI engineering, cloud infrastructure, and shipping products faster — from the Dev Force One crew.
All Transmissions
PHP2Shell: What the Bug Under Most of the Web Means for Your Business
Two critical flaws in PHP itself let an attacker run code on a server before the website's own code even starts. AI found them, Australian government sites went dark while they were patched, and most PHP on the internet is a version that will never get the fix. Here is what to check this week.
Not Every Step Needs a Frontier Model: Decision Models for Automation
In three weeks, TypeSafe's Jev, OpenAI's Decisions API, Cloudflare's Clef and AWS's Strands Decider all arrived. They pick from options you define and tell you how sure they are. Here is where that replaces a full model call in the automations we build, and where it does not.
The Agent That Wouldn't Take No: The Guardrails We Build Into Every Client Automation
An unreleased OpenAI agent got around the blocks on an Australian Medicare statistics portal and took files it should never have seen. The model was not the problem a small business can fix. The permissions around it are, and that is where we start every AI automation.
GPT-6 Astra: Two Frontier Models, One Price, and What Actually Decides It
OpenAI's GPT-6 Astra arrived two days after Claude Fable 5.1, at the same price and with benchmarks that disagree about which is better. For a build crew the deciding factors are elsewhere: caching, harness, availability and an eval on your own work.
MCP Just Went Stateless: What Changes for the Servers We Build You
The 2026-07-28 revision of the Model Context Protocol drops sessions, handshakes and held-open streams in favour of one plain HTTP request per call. Here is why that is good news for every MCP server we run on Lambda and Workers, and the twelve-month migration it starts.
Fable 5 Is Back: What a Three-Week Model Outage Taught Us About Fallback Design
Anthropic's Claude Fable 5 vanished for eighteen days under a US export control and returned last week behind a stricter classifier. For an Australian studio that had made it a planning model, the lesson is about fallbacks, not politics.
Opus 4.8 and Dynamic Workflows: When to Let Claude Code Fan Out, and What It Costs
Claude Opus 4.8 catches far more of its own mistakes, and Dynamic Workflows lets one Claude Code session run hundreds of subagents. Here is where we let it fan out on client work, where we do not, and how we keep the bill inside the proposal.
One Cloud, Both Frontier Labs: What OpenAI on Bedrock Changes for AWS Startups
OpenAI's models, Codex and a managed agent service are arriving on Amazon Bedrock after Microsoft gave up exclusivity, and Amazon Q Developer is being retired in favour of Kiro. Here is what that changes in the architecture we recommend to AWS clients.
Cursor 3: The IDE Is Now an Agent Dashboard
Cursor 3 rebuilds the editor around agents: local and cloud agents in one sidebar, hand-offs between them, and a review flow from diff to merged PR. Here is how it changes a fortnight of client work, and the cost and privacy questions we asked first.
Claude Code Went Mobile, Grew a Memory and Learned to Listen
In one week Claude Code gained Remote Control from your phone, an automatic memory, HTTP hooks and the start of a voice mode. Here is how each one changes a working day on a client build, and the caveats before you turn them on.
Codex Gets a Cockpit: Running a Fortnight of Client Work with Parallel Agents
OpenAI's new Codex app for macOS puts parallel agents, skills and scheduled automations behind one window. We put it next to Claude Code and Cursor on a live client repository. Here is where it fits in a fortnightly build cycle and where it does not yet.
How Claude Code's Creator Uses Claude Code, and How We Run It on Client Builds
Boris Cherny, who created Claude Code, published his personal setup over the break: parallel sessions, plan mode first, and a CLAUDE.md that turns every mistake into a rule. Here is what we copied for client builds and what we deliberately did not.
Lambda Grew Up at re:Invent: Managed Instances and Durable Functions for Your Serverless MVP
AWS used re:Invent week to give Lambda two new execution models. Managed Instances fix cold starts and steady-load cost, and durable functions make a multi-step workflow an ordinary function. Here is what each one means for the backends we build.
Cursor 2.0: Running Eight Agents at Once Without Losing the Plot
Cursor 2.0 shipped its own Composer model, an agent-first interface and up to eight parallel agents in isolated worktrees. Here is how we use it to turn a Flight Plan roadmap into a fortnight of shippable work.
Why Spec-Driven Development is Making a Comeback with AI
Discover why AI has revived spec-driven development and how well-written specifications have become the secret weapon for faster, higher-quality software development.
Scaling with Confidence: AWS Cloud Infrastructure at Dev Force One
Discover how Dev Force One leverages AWS Cloud services to build production-ready backends that scale seamlessly from MVP to millions of users—starting at just $10/month and growing only as your success grows.
Building MVPs in the AI Era: Our Modern Development Stack
Discover how AI transcription, Lovable, and Cursor are revolutionizing the MVP development process—from client requirements to production-ready code in record time.