Aug 25, 2026 / AI Tooling
The Claude Agent SDK is the same agent infrastructure that powers Claude Code, exposed as a library: the agent loop, built in tools for reading files and running commands and searching the web and editing code, hooks, sessions, MCP support, and subagents that run in isolated context windows and report only what matters back to an orchestrator. Two features decide whether an agent survives contact with real work. The first is context compaction, which automatically summarizes earlier messages as the context limit approaches so a long running agent does not simply run out of room. The second is the session on ClaudeSDKClient, which you can persist and resume so an agent picks up where it left off. It ships in Python and TypeScript, and from June 15, 2026 Claude subscription plans include a separate monthly Agent SDK credit. This post is about what those primitives change in how you design an agent, and where an SDK is the right tool versus a heavier orchestration framework.
Aug 25, 2026 / Headless Commerce
The choice between commercetools and Shopify Plus is not a feature checklist, it is a decision about how much of your commerce stack you want to own. commercetools is MACH infrastructure: microservices, API first, cloud native, headless, with no frontend, no opinions about your data model, and no prescriptions about checkout, priced from roughly 40,000 to more than 500,000 dollars a year before you add the headless CMS, search layer, and frontend team that a composable stack requires. Shopify Plus is the opposite bargain: an opinionated platform at roughly 2,000 to 2,500 dollars a month plus a 0.2 percent transaction fee and app spend, that ships in weeks rather than quarters. What changed in 2026 is that Shopify's Winter edition pushed hard into agentic commerce, with Storefront MCP for Hydrogen and AI discovery for headless stores, which narrows the flexibility gap that used to justify going fully composable. This post is a decision framework, not a scorecard: what each platform actually costs, how long each takes, and which revenue and complexity profile each one fits.
Aug 25, 2026 / AI Tooling
xAI moved into the agentic coding fight with Grok Build, and its opening move is price. The Grok Build 0.1 model lists at 1 dollar per million input, 0.2 dollars per million cached input, and 2 dollars per million output, which undercuts the frontier tiers most coding agents default to. Above it sits Grok 4.6 with a 500K token context window, reasoning, function calling, and structured outputs, priced at 2 dollars input and 6 dollars output per million below 200K tokens, and positioned for long running agents doing research, codebase analysis, and application development. xAI has run this playbook before: at the grok-code-fast-1 launch it partnered with seven IDE adjacent tools for a free period. This post separates what Grok Build actually ships from the noise, reads the pricing against the incumbents, and explains where a cheap, fast coding model earns its place and where it does not.
Aug 25, 2026 / Local AI Infrastructure
Ollama 0.19, released March 30, 2026, switched its Apple Silicon inference from the llama.cpp Metal backend to MLX, and on the reference test decode throughput jumped from 58 to 112 tokens per second. That is close to a doubling for the interactive part of a local workload, and it lands because MLX is purpose built for Apple's unified memory rather than ported onto it. The gain is not uniform: MLX leads by roughly 20 to 87 percent for models under 14B, then converges with llama.cpp above about 27B where memory bandwidth becomes the ceiling and the backend stops being the bottleneck. This post explains what the backend switch actually changes, where the decode gain shows up and where it disappears, how prefill and decode behave differently, and how to verify the backend and measure the numbers on your own hardware rather than trusting a headline.
Aug 24, 2026 / AI Tooling
When Cognition retired the Windsurf brand and relaunched the editor as Devin Desktop on June 2, 2026 with support for the open Agent Client Protocol, it made a bet that the editor and the agent should stop being welded together. Agent Client Protocol, first published by Zed Industries in August 2025, is a small idea with large consequences: standardize the messages between a code editor and a coding agent so that any editor can drive any agent, the same way the Language Server Protocol let any editor talk to any language server. It runs as JSON-RPC 2.0 over stdio, with the editor launching the agent as a subprocess and keeping control of the workspace, the permissions, and the user experience. This post explains how ACP works, why it is not the same thing as MCP, and what a standard editor to agent interface changes for teams building or buying agentic coding tools.
Aug 24, 2026 / AI Tooling
Computer use agents crossed a symbolic line in 2026: on the OSWorld-Verified leaderboard the top models now sit above 85 percent, past the 72.4 percent human baseline, with Qwen3.8 Max at 86.1 percent and Claude and others clustered just behind. Google shipped Computer Use in preview with Gemini 3.7 Flash, and every major lab now has an agent that can drive a graphical interface. It is tempting to read a benchmark above the human baseline as a signal that these agents are ready to run real workflows unattended. That reading is wrong, and the gap between a passing OSWorld score and a reliable production automation is where most projects fail. This post explains what an OSWorld number actually measures, why per step accuracy compounds badly over long trajectories, and how to size the reliability, cost, and latency of a computer use agent before you put it in front of a real business process.
Aug 24, 2026 / AI Tooling
Gemini 3.7 Flash shipped on August 13, 2026 as Google's cheap agentic workhorse, and on paper the generational jump over 3.6 Flash is real: FrontierCode climbs from 34.4% to 43.6% and DeepSWE from 49.0% to 65.3%, a sixteen point gain on software engineering tasks that is large for a same tier upgrade. It carries a one million token input window, decodes fast at roughly 340 tokens per second, and lands at 56 on the Artificial Analysis Intelligence Index. The number that decides whether you should move now, though, is not on the benchmark chart. It is the pricing schedule. Intro rates of $0.75 per million input and $3.75 per million output hold only through December 31, 2026, after which they double to $1.50 and $7.50. This post reads the coding and agent gains against the previous generation, works out where the speed and the window matter, and explains why the price step change should shape how you plan the migration.
Aug 24, 2026 / AI Infrastructure
Prompt caching is the rare cost lever that is both large and underused. As of 2026 all three major hosted providers discount cached input by up to 90 percent: Anthropic charges 0.10x the base input rate for a cache read, OpenAI discounts cached input automatically above a token threshold, and Google bills context caching by stored token time rather than per token. The catch is in the mechanics. Caching only ever discounts input tokens, never output, and it only pays off when a large, stable prefix is reused often enough to clear the write premium and any storage fee. Get the prefix ordering wrong, or cache a prompt that changes every call, and you pay more, not less. This post works the arithmetic that decides whether prompt caching cuts your bill, walks the provider specific pricing that changes the answer, and lays out the prompt structure that makes caching pay.
Aug 24, 2026 / Commerce Engineering
Shopify's Summer 2026 Edition made Checkout Components generally available on Shopify Plus, and it quietly redraws the build versus buy line for enterprise checkout. For years the choice was binary: accept Shopify's hosted checkout with the customization limits of checkout extensibility, or go fully headless and own the entire storefront including the checkout, taking on the PCI scope, the maintenance, and the risk that comes with it. Checkout Components is the middle path: composable, custom checkout built from Shopify's own primitives, running on Shopify's infrastructure, without going headless. For a large brand that wants a differentiated checkout but does not want to own a payment surface, that is a meaningful new option. This post reads Checkout Components against a full headless build on the axes that actually decide enterprise architecture, PCI scope, upgrade path, performance, and total cost of ownership, and gives a straight framework for choosing.
Aug 23, 2026 / AI Infrastructure
Speculative decoding usually means a small autoregressive draft model guesses one token at a time and a big model checks it. DFlash does something different: a roughly one billion parameter block diffusion drafter predicts sixteen tokens in parallel, then the full target model verifies the entire block in a single forward pass, keeping the longest correct prefix and rolling back the rest. The dflash-mlx project ports that loop to MLX on Apple Silicon with lossless greedy verification, an OpenAI compatible server, and published M5 Max benchmarks. The headline that made the rounds, a flat 4.6 times speedup on Mac, is wrong; the real numbers are model, quantization, and context specific, and they are still very good. This post explains how block diffusion drafting differs from a normal draft model, reads the actual dflash-mlx numbers on an M5 Max, and shows you how to stand it up behind an OpenAI compatible endpoint without pretending one multiplier fits every model.
Aug 23, 2026 / AI Tooling
Grok 4.6 from xAI landed on August 12, 2026 with a clear pitch: maximum intelligence per dollar. It posts 61 on the Artificial Analysis Intelligence Index, ahead of Gemini 3.7 Flash at 56, and it carries a decisive lead on knowledge work with a 1753 GDPVal-AA Elo, while charging $2 per million input tokens and $6 per million output. That is frontier tier reasoning at a price well under the premium closed models, which makes it a real candidate for production work that used to default to the most expensive option. The catch is that a low token price is not the same as a low cost per outcome, and Grok 4.6 has two properties that push in the other direction: it decodes slowly at 85.8 tokens per second and its context window is 500,000 tokens rather than the million some competitors offer. This post reads the pricing and the benchmarks together, works the cost per task math that the sticker price hides, and lays out the lanes where Grok 4.6 is the right production choice and the ones where it is not.
Aug 23, 2026 / AI Infrastructure
The question every local inference build eventually asks is whether the newer Apple Silicon is worth the money, and for the M5 Max the answer sits almost entirely in one number: memory bandwidth. The M5 Max moves to 600 GB/s from the M4 Max's 546, about a ten percent lift, and because single stream local decoding is memory bandwidth bound rather than compute bound, tokens per second tracks that number closely. Layer the M5's GPU and Neural Engine architecture improvements on top and reported throughput gains reach into the high twenties percent on some workloads, though the real figure depends heavily on the model, the quantization, and the context length. This post explains why decode throughput follows memory bandwidth, reads the M5 Max against the M4 Max on the numbers that matter for running models locally, and gives you a straight answer on when the generational jump justifies the upgrade and when your M4 Max is fine.