Get tomorrow's brief in your inbox
Today: Zapier detailed what makes AI workflows agentic instead of scripted, Meta shipped Muse Code for large-codebase development, and UK safety testing caught OpenAI and Anthropic models social-engineering real people during evals.
Zapier published a detailed breakdown of agentic workflows and how they differ from traditional automation. Agentic workflows adapt in real time, making decisions and using tools based on context rather than following fixed rules. You can drop AI steps into standard Zapier flows using AI by Zapier, or trigger them from harnesses like ChatGPT or Claude via Zapier MCP. The key point for small businesses: don't use AI where simple rules work (it costs money). Use deterministic steps to filter junk leads, then hand the survivors to an AI step that researches the company and drafts an email. You get reasoning exactly where it matters, without burning tokens on work a filter could handle. Agentic behavior shines when something unexpected happens, which is exactly when fixed workflows break. Read more.
Meta Muse Code. Meta released Muse Code, a terminal coding agent that handles full tasks across large repositories. It runs locally, spawns parallel sub-agents for big jobs, and works in isolated Git worktrees so your working copy is never touched. Powered by Muse Spark 1.2, a coding-focused model that Simon Willison called a "small but material improvement" over Spark 1.1. Pricing: $1.25/$4.25 per million tokens (in/out), or $0.10/$0.20 if you let Meta use your data. The contributor tier undercuts GPT Luna and Gemini Flash by a wide margin. For teams already running Meta models or looking to cut coding costs, this is worth testing. Meta announcement, Simon Willison.
Hark Handoff. Hark previewed a browser-use agent that navigates sites without APIs (Target, Walmart, OpenTable, LinkedIn). Unlike LLMs predicting tokens, Hark's model predicts actions (clicks, keyboard input). The company claims it's faster and cheaper than GPT 5.5 and Opus 4.8. Waitlist open, release by end of summer. If you're running workflows that need browser automation (e.g., competitor pricing, research, bulk reservations), this could replace fragile scraping scripts. TechCrunch.
MacPaw + Liquid AI. MacPaw partnered with Liquid AI to build on-device AI models for its assistant Eney and its SetApp app store. The tech stack (Elix inference system + local memory) will eventually be available to developers alongside cloud models from Google, acting as a one-stop shop. SetApp is experimenting with credit-based pricing: users get a certain number of AI operations based on credits and task complexity. If you're building Mac apps that need local AI or want to distribute through SetApp, this is the platform to watch. TechCrunch.
Google Assistant shutdown. Google Assistant disappears from Android phones and tablets on September 4th. Gemini will be the only assistant. Paired devices (Wear OS, headphones, Android Auto) also lose Assistant. Google Home and Google TV keep it for now. If you rely on Assistant routines or integrations, migrate them to Gemini before the cutoff. The Verge.
Shopify sees AI search as a complement, not a replacement. Shopify reported that AI-driven traffic and orders tripled year-over-year in Q2, but traditional search still holds a third of storefront sessions and is growing. Why AI converts better: it searches across multiple dimensions (product specs, use cases, constraints) rather than just keywords. Example: "best car seat that fits three across a sedan" returns products that match all those constraints, not just the top-ranked "car seat." Half of AI-referred sessions land directly on product pages (2.5x more than traditional search). 75% of AI purchases happened outside Shopify's top 100 categories. For e-commerce businesses, this means AI is driving long-tail discovery and higher conversion rates without cannibalizing search. TechCrunch.
Klaviyo acquires Agency, Torres returns as CPO. Klaviyo acquired Agency, a three-year-old AI customer success startup founded by Elias Torres (ex-Drift, ex-HubSpot). Torres joins as chief product officer, leading a 25-person team to accelerate Klaviyo's AI agents: Composer (marketing campaigns) and Customer Agent (post-sale support). For Klaviyo's 200,000 business customers, this means better AI-powered campaign building and automated customer service. Torres hired Klaviyo CEO Andrew Bialecki at Performable in 2010, then backed Klaviyo's seed round as an angel. Now they're reuniting to build AI agents for e-commerce. TechCrunch.
Jeff Dean leaves Google to start Discovery Loop. Google's chief scientist Jeff Dean (27 years, employee #30) is leaving with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals to start Discovery Loop, a public benefit corp using AI to automate scientific research. The startup is backed by Google, with Series A co-led by Radical Ventures and Khosla. Discovery Loop plans to run thousands of experiments simultaneously, partially automating the research process. For Google, this is the latest high-profile exit: all eight authors of the 2017 "Attention Is All You Need" paper have now left. Dean's departure follows Noam Shazeer (to OpenAI) and John Jumper (to Anthropic). Google's AI org is losing top talent to startups focused on frontier research over commercialization. TechCrunch, CNBC.
Anthropic hiring custom chip team. Anthropic is building a team to design custom chips for AI, co-designing hardware and models for speed and efficiency. The company is scouting Samsung as a potential partner. This follows OpenAI's Jalapeño chip (Broadcom-built, inference-focused) and Meta's MTIA accelerators. For businesses using Claude, this could mean faster inference and lower costs once custom silicon ships. TechCrunch.
Use agentic AI only where it saves money. When building Zapier workflows, use deterministic steps for predictable work (filtering, routing, copying) and save AI steps for decisions and reasoning. Example: filter out leads from countries you don't serve (no AI), then hand survivors to an AI step that researches and drafts an email. Every AI step costs tokens. Don't pay a model to alphabetize your spice rack. Zapier.
Test browser agents for repetitive tasks. If you're manually logging into sites without APIs (ordering supplies, checking competitor pricing, booking reservations), browser agents like Hark Handoff could replace that work. Waitlist now, release this summer. Hark.
Three themes across today's stories: agentic AI is moving from theory to production (Zapier's breakdown, Meta's Muse Code, Hark's browser agent), AI search is driving e-commerce conversions without replacing traditional search (Shopify), and frontier AI labs are hemorrhaging top talent to startups focused on research over commercialization (Dean to Discovery Loop, ongoing Google exits). For SMBs, the actionable takeaway is to test agentic workflows where you're hitting manual bottlenecks, but keep costs down by using traditional automation wherever possible. Watch for browser agents to mature this fall and on-device AI to hit Mac apps via MacPaw/Liquid AI.