Get tomorrow's brief in your inbox
Today: Zapier launched multi-agent orchestration with MCP to coordinate specialized AI teams across 9,000 apps. Google debuted Gemini Omni with natural language video editing and Gemini 3.5 Flash for production-grade agents. Meanwhile, coders refuse to work without AI despite evidence it generates 1.7x more bugs than human code.
Multi-agent systems are moving from theory to production, and Zapier just made them accessible to non-developers. Their new MCP (Model Context Protocol) server lets you build specialized AI agents that coordinate across your entire tech stack without writing code. Instead of one overloaded agent trying to handle everything (and failing), you create focused agents for specific tasks: one qualifies leads, another researches companies, another drafts emails. Connect them through Zapier's 9,000+ app integrations, and they work as a team.
This matters because single-agent systems hit a wall fast. When you ask one AI to juggle multiple tools and decisions, its reasoning degrades. The fix is orchestration: divide the work, scope each agent's access tightly, and let them pass context between tasks. Zapier's approach uses OAuth-managed connections and SOC 2 infrastructure, so your credentials never touch the models. For small businesses running customer support, lead qualification, or content workflows, this is the difference between "AI could help someday" and "AI is handling this today." The learning curve just dropped from weeks of coding to hours of workflow mapping. (Zapier, Zapier orchestration)
Google Gemini Omni turns videos into editable assets through conversation. Upload a clip, then ask it to "make the sculpture out of bubbles" or "transport the violinist to this environment." Each instruction builds on the last while keeping characters and physics consistent. This isn't generating from scratch; it's editing what you filmed with natural language. For product demos, training content, or social media, you can iterate faster than switching between video tools. Available through Google AI Studio. (Google)
Gemini 3.5 Flash ships with agent-grade reliability. Google's new model balances flagship intelligence with Flash-level speed, making it viable for long-horizon tasks like multi-step workflows and coding projects. When paired with Antigravity (Google's agent harness), it executes supervised multi-step work without degrading. Now live as the default in Gemini app, AI Mode in Search, and AI Studio. If you've been waiting for agents that don't need constant hand-holding, this is the generation to test. (Google)
Android XR intelligent eyewear hits shelves this fall. Google partnered with Samsung and Qualcomm to build audio glasses (this fall) and display glasses (later). Audio versions from Warby Parker and Gentle Monster give you hands-free Gemini assistance: ask questions, get directions, manage tasks while staying heads-up. For restaurant managers coordinating staff or retail owners checking inventory, this could replace pulling out your phone fifty times a day. Display glasses add visual data overlays. Privacy and cost will determine adoption, but the use case is clear for high-motion environments. (Small Business Trends)
OpenAI Codex (GPT-5.5) is converting customer requests into working code in minutes. Braintrust engineers now take feature requests, paste them into Codex, and show customers preview branches before the meeting ends. Half the team moved to Codex in one month. The shift isn't just speed; it's faster feedback loops with customers. Instead of backlog purgatory, requests get prototyped live. If you build software products, this changes scoping conversations: you can show working ideas, not mockups. (OpenAI)
Boston Children's Hospital embedded AI as infrastructure, not experiments, saving 60,000 hours and redeploying $7M in labor costs across clinical and operational workflows. The hospital built an enterprise ChatGPT layer used by over one-third of employees daily. Supply chain teams automate invoice routing. Surgical scheduling uses AI to analyze notes and estimate patient acuity, filling OR time further in advance. Clinical teams diagnosed 40+ rare conditions that had gone unresolved by synthesizing fragmented genetic data and medical literature faster than humans could. The lesson: treat AI as shared infrastructure with governance built in, not one-off tool experiments. Deploy where repetitive work meets high volume, then measure operational impact in hours and dollars. (OpenAI)
AI coding productivity is a mirage. Developers now refuse to work without AI (METR couldn't recruit participants for a study because devs "do not wish to work without AI"), yet the evidence shows problems. CodeRabbit reports AI generates 1.7x more bugs than human code. Companies spend 44% of tokens fixing AI-generated bugs (per Entelligence AI's analysis). Amazon shut down its token-tracking leaderboard after employees gamed it without productivity gains. Uber blew its 2026 AI budget in four months with no measurable project increase. The issue: AI writes code fast, but you spend the savings (and more) debugging, steering, and maintaining it. Treat AI as a draft generator, not a replacement for thinking through the problem. (TechCrunch)
Zapier vs. Workato for enterprise agents: control vs. speed. Workato requires IT to design agent pathways and approve connections, making it ideal for regulated industries (HIPAA compliance, on-premise deployment). Zapier lets anyone build agents within IT-set guardrails, cutting time-to-value. Workato fits deep enterprise platform integrations (Salesforce, SAP). Zapier covers niche tools and lets teams solve their own problems. If your culture is "IT controls production," go Workato. If you want distributed innovation with governance, Zapier wins. (Zapier)
Shift AI will clean your home for free in exchange for training footage. The startup sends cleaners wearing camera-equipped hats to record scrubbing, mopping, vacuuming. The value of robot training data exceeds cleaning costs. Starting in NYC, expanding to SF, London, Zurich, Munich. Privacy policy: faces, names, screens blurred before AI training. Free cleanings are "limited time," but this is the model now - physical-world data is expensive, so companies get creative. If you're in an eligible city and comfortable with cameras, free cleaning. Otherwise, recognize this trend: your actions are becoming training data. (The Verge, The Verge coverage)
Snowflake's $6B AWS deal + raised guidance drove 50% stock jump, marking the best week for software stocks since 2001. Data platform customers are deploying AI workloads faster, and Snowflake sits underneath as "picks and shovels" infrastructure. Okta surged 30% on earnings, reporting agentic AI forces businesses to invest in identity security (bot armies need defenses). The "SaaSpocalypse" narrative cooled. Software that enables AI (not gets replaced by it) is winning. (CNBC, Snowflake earnings)
XCENA raised $135M at $570M valuation betting memory, not compute, is AI's bottleneck. Their MX1 chip places compute near DRAM, handling data operations (preprocessing, KV cache management) within memory modules instead of round-tripping to GPUs. If it scales, this cuts inference costs dramatically. Mass production scheduled end of 2026, revenue in 2027. Target customers: hyperscalers spending billions on AI infrastructure. This is a bet that inference efficiency matters more than raw GPU power. (TechCrunch)
Groq reportedly raising $650M to grow inference cloud after $20B Nvidia not-acqui-hire. In December, Groq licensed hardware tech to Nvidia and some execs departed for cash. Now existing investors (Disruptive, Infinitium have committed to fill the round) are funding Groq's pivot to inference-focused neocloud hosting. Inference processing (after a prompt) is a bigger current need than training. Groq's homegrown AI chips target that workload. (TechCrunch)
Cognition raised $1B at $26B valuation for Devin coding agent, with CEO Scott Wu insisting the goal isn't replacing programmers. Devin commits 89% of Cognition's own code, mostly handling long-tail maintenance (updating old software, platform migrations). Wu positions it as freeing developers "from toil" to focus on creation. Reality check: Devin works at "junior to mid-level" depending on task, and still needs human oversight. The "self-driving software" framing is ambitious; the current state is "automated janitorial work for codebases." (TechCrunch)
Build a lead qualification multi-agent workflow. Set up three Zapier agents: (1) qualifies incoming form leads by checking company size and budget fit, (2) researches the company's recent news and tech stack, (3) drafts personalized outreach email. Connect them sequentially so context flows automatically. This takes manual research out of your sales funnel and speeds response time from hours to minutes.
Use webhooks instead of polling for real-time automation. If you're checking apps repeatedly for new data (polling), switch to webhooks: the app sends data to your automation URL when events happen. Faster, less work, no repeated requests. Most modern apps support webhook triggers. Check your automation platform's webhook documentation and set up a listener URL. (Zapier webhooks guide)
Three patterns stand out today: (1) orchestration beats power - multi-agent systems with tight tool scopes outperform single overpowered agents, (2) infrastructure enablers win - Snowflake, Okta, memory-centric chips are thriving while pure AI apps face "SaaSpocalypse" fears, (3) productivity theater is collapsing - token usage doesn't equal output, and the AI code maintenance bill is coming due. Watch for: more MCP-compatible tools launching, increasing scrutiny on AI ROI claims, and backlash against forced AI features (DuckDuckGo installs climbing as users flee Google's AI search). The honeymoon phase is ending; now comes the real work of deploying AI where it actually pays off.