Quick Hits

• @ClaudeDevs announced it is open-sourcing Claude Commerce Agents, a forkable blueprint with runnable reference implementations across retail, travel, telecom, and entertainment. Per @MaxForAI's detailed breakdown, Anthropic claims some retail clients saw cart value rise up to 35% and purchase completion up to 60%, with Shopify and Priceline already building shopping experiences on Claude.
• Shopify's @tobi says the company's ML team fine-tuned a 0.8B model that beats GPT 5.6-sol xhigh on "this very specialized task," crediting a "self improving recursive flywheel." @Ales_is_offline reads it as a signal that in-house small models, and the ML teams to run them, are becoming a must-have.
• @guanlan's FrontierHarness Eval ran the same model on the same tasks across nine harnesses, including Claude Code, Codex, Pi, Exo, and DeepSeek Harness: 360 runs and 2 billion tokens later, pass rates ranged from 50% to 67% and cost per pass from $1.05 to $18.34.
• @finkd says Meta's Muse Spark 1.3 is rolling out with "frontier performance almost too cheap to meter," calling it the company's biggest jump yet on coding and agentic work, with open weights releases "coming soon." @MatthewBerman's reaction: "Ok I see you Meta."
• Briefly: @jimfarley98 revealed Ford's 2027 midsize electric truck, the Fathom, during a desert-testing detour through Las Vegas; @QwenDevs introduced zg, a local-first search tool going "beyond keywords" for humans and agents; @0xHvdes says @0xfJuan's "Everywhere Overnight" startup playbook is worth the $1-3k advisors charge; and @i2cjak cheered @amembedded's $63 STMicro VL53L9CX LiDAR breakout (8.8m ranging, 100Hz) as proof "American embedded is goated."

Anthropic Bets on Being E--Commerce's Middle Layer

The clearest strategic signal in today's feed: Anthropic wants to power commerce agents, not compete with merchants. Per the @ClaudeDevs announcement and @MaxForAI's summary, the open-source blueprint splits into two agents. A Shopping Agent embeds in a merchant's app or site to handle search, comparison, personalized recommendations, cart building, order status, and after-sales. A Merchant Agent serves internal staff with sales analysis, inventory, pricing, and campaign generation, with any mutation to catalog or prices routed through staging for human approval.

The architecture is deliberately plain: one Claude model running an agent loop with tools, skills, memory, guardrails, and evals, calling the merchant's existing search, cart, inventory, and pricing systems. No multi-agent routing. Even UI elements like product cards, itineraries, and seat maps are exposed as structured tools rather than generated text. There is also a Claude Code plugin, so a repo can generate a custom commerce agent via a command like /scaffold-commerce-agent.

Notably, Anthropic says it will not build marketplace, payment, or ad layers; catalog, checkout, and customer relationships stay with the merchant. The usage numbers are Anthropic's own claims relayed through social posts and have not been independently verified.

Tiny Fine-Tunes and Token Thrift

The counterpoint to frontier-model hype ran through several posts. @tobi's claim that a 0.8B fine-tune beats GPT 5.6-sol xhigh on a narrow task came with a caveat worth keeping: he attributes it to Shopify's self-improving flywheel, not small models being universally better. @Ales_is_offline frames the economics: companies may stop paying premium prices for general models on tasks an internal small model handles.

The cost theme extends to tooling. @dr_cintas posts four concrete steps to stop Fable 5.1 from burning weekly credits: set /effort to low, run /claude-api cost-optimize, run /claude-api prompt-audit to clear stale junk, and migrate old model IDs to claude-fable-5-1. And @0xIlyy reports, secondhand, that one tinkerer pushed GLM 5.3 Flash on two RTX PRO 6000 Blackwell cards from 200 to 1,000 tokens per second, with 20-50% gains arriving every few hours. His conclusion: "whatever software you use, it's nowhere near to being optimized enough."

Harnesses, Factories, and Triggers: The Plumbing Takes Over

If models are commoditizing, today's posts put the differentiation in the scaffolding. @guanlan's FrontierHarness Eval makes the case with numbers: identical model, tasks, and runtime, yet pass rates span 17 points and cost per pass varies more than 17x depending on the harness. @GoogleCloudTech published a guide on the same problem from the cost side, explaining that agent harnesses manage runtime state, sandbox execution, and prompt preparation, and that re-sending static context every turn multiplies token spend; context caching is the fix.

@stencil_labs argues the deeper point in a teardown of what OMP got wrong and OMP² changes, claiming "a while loop around a fetch" is really "systems software in a trenchcoat"; @viathebrink calls it "an actual goldmine for sota harness engineering." The factory metaphor keeps spreading too: @augmentcode shared co-founder @igoro's look at software factories being built at Cloudflare, Vercel, Uber, and LaunchDarkly, ending on the question of whether agents will eventually be prompting us. @kdaigle flagged @burkeholland's confession of having "accidentally made an agent factory" and the surprising things agents do while coordinating.

On the consumer side, @ericzakariasson (via @RayFernando1337) says Grok Bot routines can now be triggered by webhooks, from a specific WhatsApp contact messaging you to a Notion page changing, a temperature threshold, a desk button, a GitHub action finishing, or a server error spike. @RayFernando1337's framing: "Zapier but on steroids and with plain English." @adiix_official goes further, documenting how he fed all 18 rules from a "3-page SpaceXAI manual" into Grok Bot to build a Chief-plus-specialists setup with a shared /workspace, three gates (source, evidence, action), and humans approving only money, publishing, deletion, and signatures.

Code Review Meets Slop, and the Curriculum Rewrites Itself

@martinfowler linked Rachel's Ramblings asking "Maybe We Shouldn't Be Reviewing All This Code," arguing AI may not have broken review so much as exposed that review was being used to solve the wrong problems. @henrytdowling's slop-fighting list is blunt: agents give unit tests too much authority, and bad tests are a huge slop source. His remedies include @dillon_mulroy's antislop oxlint rules, @_lopopolo's "useless test deleter," telling the agent to "delete as much as possible," and adversarial goals on LOC or Halstead complexity.

@0xblacklight, crediting @humanlayer_dev, adds the compounding worry: models improve at working in messy codebases, but they improve even faster at creating mess, and that debt "compounds superlinearly on large teams." Quoted @Steve_Yegge puts it more personally: every model "will eventually build systems that they can no longer understand," and Fable 5 flailed on him for a week until 5.1 arrived. His advice is to keep "an iron grip on system size."

Tooling is responding. @github shipped a repeatable --attach flag in GitHub CLI for inline images and video in issues and PRs, which @OhansEmmanuel combines with pr-lens for animated architecture diagrams; @mattpocockuk praises @dexhorthy's /show-me skill as "a toolbox of nice ways to look at code" that makes PR descriptions readable. On education, @mihail_eric's updated Stanford course throws out 85% of last year's material for agent skills, context engineering, MCP portals, and agentic review, with students shipping PRs to partner OSS repos and all classes public from 9/22; @nummanali calls the free resources a gift. @DeRonin_ posts a 6-month robotics engineering resource list, calling robotics "the least crowded high-value skill in tech right now." And on output craft, @addyosmani flags Anthropic's official de-flavoring prompt for stripping mannered prose from Fable 5.1, @wp confirms @lennysan's roundup of @anshuc's eight creativity techniques (seed strings, ambitious prompts, subagent loops, image and video generation, hand-rewritten copy) "actually works" for divergent design, and @alex_prompter keeps circulating prompt shares.

Practical Takeaway

Before you pay for a bigger model, benchmark the scaffolding and the task. If you have one narrow, high-volume workflow, @tobi's post is a reason to prototype a small fine-tune against your frontier baseline. If your costs are the problem, @guanlan's eval suggests running identical tasks across two harnesses first, since the spread there dwarfed model-to-model differences, and @dr_cintas's caching and audit steps apply immediately. Whatever you pick, @Steve_Yegge's warning stands: set explicit size and slop limits on agent-written systems now, because the compounding mess is harder to undo than the savings are to capture.