Technical deep-dives, AI strategy, and engineering perspectives from the team building autonomous systems.
Both are Postgres. That is where the similarity ends. One is a storage engine rebuilt for elasticity, the other is an application backend with Postgres at the center, and choosing wrong costs you a rewrite.
The AI coding agent field in August 2026 has stopped being about who has the smartest model and started being about who fits the way developers actually work. The adoption data tells the story: awareness of Codex climbed from 27 percent of developers early in the year to 65 percent by mid year, while GitHub Copilot's share slid from 29 percent a year ago to 21 percent, according to JetBrains research. On raw capability the top agents are bunched together, with Terminal-Bench 2.1 putting GPT-5.6 Sol, Claude Opus 5, and Grok 4.6 all within roughly a point of each other. When the models are this close, the differentiator is workflow shape, not benchmark position. This post lays out where adoption is heading, what the benchmarks do and do not tell you, and the multi tool stack most working developers have converged on, because almost nobody uses just one agent anymore.
GPT-5.6 Sol is the flagship tier of OpenAI's GPT-5.6 family, released on July 9, 2026 alongside Terra and Luna. It ships a 1,050,000 token context window and, after OpenAI cut prices on August 22, sits near 4 dollars per million input tokens and 20 dollars per million output tokens. On Terminal-Bench 2.1, the agentic coding benchmark, it lands at the very top of the public leaderboards, though the exact number swings with the harness and the effort setting used. This post skips the launch-day cheerleading and treats Sol as a procurement decision: what it costs in production, what the benchmark numbers actually say once you account for effort tiers, and the specific workloads where it beats the cheaper Terra and Luna tiers or a rival like Claude Opus 5. If you are choosing a default model for an agent, the interesting question is not the headline score, it is the cost per solved task.
A model called OX Alpha showed up as an anonymous frontier system, free to use through August 27, 2026, carrying a 1 million token context, multimodal input, and a reported score near 80 percent on DeepSWE, an agentic software engineering benchmark. Technical analysis points to Zhipu GLM-5.x lineage behind the mask. Stealth launches like this, where a lab floats a strong model under a codename on a router or an arena before it admits what it is, are becoming a standard move in 2026, and they are worth understanding because they are how you get an early, unbranded read on the frontier. This post explains what OX Alpha is, what a DeepSWE score around 80 percent actually means for real software work, and why the stealth pattern exists at all. The number is the least interesting part; the reason a lab would ship it anonymously is the part that tells you where the field is.
As part of its August 2026 platform updates, Shopify is ending Delivered Duty Unpaid on Managed Markets, its cross border commerce service. In plain terms, stores can no longer let the international customer settle import duties with the carrier on delivery; the duty has to be calculated and collected at checkout instead, the model known as Delivered Duty Paid. That sounds like a billing footnote and it is anything but. The point of sale where a shopper learns the true landed cost of an order is the single biggest lever on international conversion and on the surprise fee refusals that generate returns and chargebacks. This post explains what the change actually is, why Shopify is making it, and what enterprise and headless stores should do about checkout, catalog data, and their merchant of record posture before the shift bites. Cross border is where commerce infrastructure decisions become finance decisions, and this is one of them.
For most of its life vLLM was a Linux and NVIDIA story, and running it on a Mac meant emulation, partial builds, or giving up. In 2026 that changed: vLLM now has two competing Apple Silicon ports. vllm-metal is a Metal GPU backend that unifies MLX and PyTorch under one compute pathway and has landed in Docker Model Runner on macOS. vllm-mlx is built directly on Apple's MLX array framework and, on an M4 Max, has posted throughput up to 525 tokens per second on small text models, exceeding llama.cpp by 21 to 87 percent across the configurations tested. Two ports with the same goal and different foundations means the interesting question is not whether vLLM runs on a Mac, it is which port to reach for and for what. This post compares them on throughput, run to run stability, and fit, and is honest about the case where the right answer is still to develop on the Mac and serve on Linux.
In August 2026 Thomson Reuters launched Thomson, a proprietary model it trained in house on an open-weight foundation, reportedly the most recent base being Qwen 3.5, combined with decades of Westlaw, Practical Law, Checkpoint, and Reuters content. The reported numbers are the interesting part: a total investment near 40 million dollars covering talent and compute, but a final training run of only about 450,000 dollars, with early evaluations placing it comparable to leading frontier models on its target tasks. That gap between 40 million and 450,000 is the whole story of the build versus buy decision, and most teams read it wrong. This post is a decision framework, not a cheerleading piece: what you are actually buying when you build a domain model on open weights, what the real costs are beyond the headline training run, and the specific conditions under which owning a model beats calling a frontier API.
For most of its short life, MCP authorization meant a bearer token in a header and a lot of hope. That is changing. The 2026-07-28 specification release and the roadmap updated on 2026-08-22 push MCP toward a real enterprise security model: proof of possession bound tokens through DPoP, short lived credentials through Workload Identity Federation, and a working group defining how an agent presents its own identity or a user delegated one. This post reads those changes the way an operator running an MCP server in production would. What is actually landing, what problem each piece solves, and what you should do now versus wait on. The short version: static keys on a remote MCP server are the liability to retire first, and the new primitives give you a path off them that does not depend on every client implementing the same thing on day one.
MCP historically had two ways to talk: stdio for a local server on your machine and HTTP for a remote one, each with its own quirks. The 2026 transport work collapses that split. Streamable HTTP is becoming the single binding, with remote servers holding a persistent connection over Server-Sent Events instead of the older polling approach, and the roadmap extends the same HTTP model to cover local servers speaking Streamable HTTP over stdio. That sounds like plumbing, and it is, but plumbing is exactly what decides whether a remote MCP server is production grade or a demo. This post explains what Streamable HTTP actually changes, why persistent SSE connections beat polling for agent workloads, and what an operator running remote MCP servers should do about it, in plain terms rather than spec language.
Shopify's move to Checkout Extensibility is not a UI refresh, it is the deprecation of the entire old customization surface: checkout.liquid, Additional Scripts, and Shopify Scripts are being replaced by Checkout UI Extensions, Shopify Pixels, Shopify Functions, and the Checkout Editor. The deadlines are real and staggered. Shopify Scripts stop working on June 30, 2026. Plus stores began hitting hard blocks on legacy checkout.liquid customizations around August 13, 2026, affecting roughly 5,200 merchants, and non-Plus stores face an August 26, 2026 upgrade deadline after which legacy customizations are removed automatically. The dangerous part is the failure mode: after an auto upgrade, tracking pixels, tag managers, affiliate scripts, and post purchase apps can stop working with no visible storefront error. This is a decision and execution guide, not a countdown clock: what changed, what breaks silently, and how to migrate without losing conversion tracking.
Shopify's Winter 2026 edition pushed hard into intelligence and automation, and the clearest expression of that bet is Sidekick, the admin assistant that now sits inside the merchant dashboard. The pitch is simple: instead of clicking through menus to build a discount, segment customers, or read your analytics, you ask Sidekick in plain language and it does the work in the admin. This post separates what Sidekick actually ships in the Winter 2026 cycle from the marketing, reads it as an operator would, and explains where an in admin agent genuinely saves time and where it is still faster to do the task yourself. Sidekick is one of more than 150 updates in the release, and it sits alongside Agentic Storefronts and a set of checkout and analytics changes that together signal where Shopify thinks commerce is heading.
Google relaunched Antigravity at I/O on May 19, 2026 as a five surface platform: a desktop app, a CLI, an SDK, a Managed Agents API, and an Enterprise Agent Platform, all built around running several agents at once with scheduled background tasks and Gemini 3.5 Flash as the default model. Cursor made the same conceptual move, reworking its core loop from a single composer to multiple agents running in parallel with a dedicated Agents view and cloud agent environments that run on isolated copies of your repo. The two products now rhyme, which makes the choice less about who has agents and more about how each one exposes orchestration, where the work executes, and what you pay to run it. This post compares the two on architecture, control surface, execution model, and pricing, and explains which team profile each one actually fits.