Home
Mar 21, 2026
View All

Notion Launches Custom Agents & Google Brings Full-Stack Vibe Coding to AI Studio

One-Line Summary#

Notion ships Custom Agents as a first-party product alongside a detailed security writeup, Google upgrades AI Studio with a full-stack vibe coding experience powered by a new Antigravity agent, Cursor launches Composer 2 with transparent per-token pricing, and Wispr Flow expands to Android with free unlimited dictation — together signaling that autonomous agents and AI-native workflows are graduating from developer tools into mainstream productivity software.

Notion — Custom Agents: Autonomous AI Teammates for Workflow Automation#

Source: https://www.notion.so/releases/2026-03-20 | https://www.notion.com/blog/introducing-custom-agents | https://www.notion.com/blog/how-we-built-security-into-custom-agents | https://www.notion.com/blog/how-notion-uses-custom-agents Credibility: High (first-party product launch with multiple companion posts including security architecture and internal usage documentation)

What happened: Notion shipped Custom Agents on March 20 — described as "AI teammates that handle entire workflows autonomously." The launch is substantial: Notion published four companion pieces simultaneously, covering the product announcement, the security model built to support it, how Notion itself uses Custom Agents internally, and a customer case study. Custom Agents are not chatbots or suggestion tools — they're persistent, workflow-level agents that users configure to handle repetitive tasks end-to-end inside Notion workspaces.

Key capabilities:

  • Agents handle entire workflows autonomously inside Notion — not single-step actions, but multi-step sequences (e.g., intake a Slack message → create a Notion page → assign action items → update a project tracker)
  • Designed to "take the repetitive work off your hands so you can focus on the work that needs you" — positioning is workflow automation, not AI chat assistance
  • Security model built from the ground up for multi-tenant agent execution: the companion post from Sean Keenan and Jacob Brackett details how Notion isolated agent permissions, scoped data access, and stress-tested the system at scale before GA
  • Notion's own marketing team uses Custom Agents to automate repetitive busywork — the internal use case post provides a concrete example that validates production readiness
  • Available as of March 20, 2026

Why it matters for PMs: This is a significant category signal. Notion — a product known for structured productivity and collaboration — is shipping agents as a first-class product, not a beta experiment. The four-post launch (announcement + security + internal use + customer case study) is a mature launch pattern that signals Notion believes agents are ready for mainstream teams, not just early adopters. For PMs evaluating where to build agent automation for knowledge workers, this raises the bar: Notion users will now expect agent-level automation from their productivity stack. The security companion post is especially useful — it documents the permission scoping and isolation architecture Notion built, which is a reference model for any PM thinking about trust and safety in agent products. Watch whether Notion's agent adoption metrics (if shared) confirm that knowledge workers outside engineering are actually using agents in production.

Critical questions:

  • What's the pricing model for Custom Agents — are they included in existing AI plans, or a separate add-on that changes the buy vs. build calculus for teams already paying for Notion AI?
  • How do Custom Agents handle failure recovery — if an agent takes a wrong action mid-workflow, can users inspect, undo, or override without losing work?
  • What integrations does the agent support beyond Notion itself — can it read from Slack, write to Jira, or update Google Sheets, or is it Notion-only?
  • The security post says agents were tested "at scale" before GA — what does that mean in practice, and what failure modes were discovered and resolved?

Action you could take today: If your team uses Notion, identify one workflow where someone is manually copying information between tools or pages more than twice a week — that's a Custom Agent candidate. Configure a basic agent on that workflow and measure: does it complete the task without error, and does the person it replaces actually trust the output? That trust measurement is the real metric, not task completion rate.

Google AI Studio — Full-Stack Vibe Coding with the Antigravity Agent#

Source: https://blog.google/innovation-and-ai/technology/developers-tools/full-stack-vibe-coding-google-ai-studio/ Credibility: High (first-party Google blog post co-authored by Ammaar Reshi, Product and Design Lead at Google AI Studio)

What happened: Google shipped a "completely upgraded vibe coding experience" in AI Studio, powered by a new agent called Antigravity. The post is co-authored by Ammaar Reshi (tracked: Product and Design Lead, Google AI Studio) and Kat Kampf. The upgrade enables users to build production-ready, multiplayer, and full-stack apps from a single prompt — not just frontend mockups or static prototypes. Antigravity is positioned as a full-stack coding agent capable of handling backend logic, database connections, and deployment configuration, not just UI generation.

Key capabilities:

  • Single-prompt to full-stack app: users describe what they want and Antigravity handles frontend, backend, and deployment configuration
  • Multiplayer support: apps built in AI Studio can support real-time multi-user interaction out of the box
  • Production-ready output: the framing is "production-ready apps," not prototypes — a deliberate contrast with earlier vibe coding tools that produced demos
  • Integrated into Google AI Studio, which means Gemini model access and Google Cloud deployment pathways are native
  • Antigravity is the agent's name — a named agent inside a product is a product design choice worth noting (it creates a memorable identity for the capability, similar to how Windsurf named Cascade)

Why it matters for PMs: Google is now competing directly in the vibe coding / AI app builder space alongside Lovable, Bolt, Replit, and Cursor. The "full-stack and multiplayer" positioning is a specific product bet: rather than competing on UI polish or ease of use, Google is claiming capability depth (backend + multi-user) as the differentiator. For PMs tracking the build-vs-buy calculus for internal tooling, this is significant — Google AI Studio is free (or low-cost via API), which undercuts paid tools like Lovable ($200M ARR) on price. The Reshi co-authorship also signals this is a product-led initiative, not a research demo: when a PM and design lead co-writes the launch post, it's a product team taking ownership, not a research team releasing a preview.

Critical questions:

  • What does "production-ready" actually mean in Antigravity's output — does it include auth, error handling, and rate limiting, or just functional code that runs?
  • How does Google handle data privacy for apps built in AI Studio — if users build apps that handle user data, what are the terms?
  • Is Antigravity available globally or in limited regions, and does it require a paid Google Cloud account for production deployment?
  • How does Antigravity compare to Lovable or Bolt on complex apps — does "full-stack" hold up when requirements get non-trivial?

Action you could take today: Pick one internal tool your team has been meaning to build (a dashboard, a simple CRM, a lightweight status tracker) and try building it with Antigravity in AI Studio. Time how long it takes to get to something usable, and note where the agent fails or requires manual intervention. Compare that experience against Lovable or Replit if you've tried those — the delta in capability and friction is the real product signal.

Cursor — Composer 2 with Per-Token Pricing#

Source: https://cursor.com/changelog/composer-2 Credibility: High (first-party changelog entry, March 19, 2026)

What happened: Cursor shipped Composer 2 on March 19, described as "frontier-level coding performance with strong results on challenging coding tasks." The notable product decision: Composer 2 is priced on a per-token basis rather than included in the flat monthly plan. Two pricing tiers: Standard at $0.50/M input and $2.50/M output tokens; Fast (default) at $1.50/M input and $7.50/M output tokens. This is a meaningful pricing architecture change — Cursor is introducing consumption-based pricing for its highest-capability model, separate from the monthly subscription.

Key technical details:

  • Two tiers: Standard (lower cost, presumably higher latency or slower queue) and Fast (default, 3x input / 3x output cost)
  • "Frontier-level coding performance" — likely refers to a new or upgraded underlying model, positioning Composer 2 above existing Cursor models
  • Per-token pricing makes Composer 2 usage directly measurable and attributable — useful for teams that need to budget AI coding costs per developer
  • Priced for power use cases: at Fast tier, a developer doing heavy code generation could spend meaningful dollars per session

Why it matters for PMs: Cursor is threading a needle here: keep the monthly flat rate accessible for standard use, while making the highest-capability model consumption-priced for teams that need it. This is the same pricing architecture emerging across the sector (Windsurf's Max plan, OpenAI's reasoning model tiers). For PMs building developer tools or managing engineering team tooling budgets, this introduces a new cost variable: AI coding tool spend is no longer purely predictable from headcount. The Fast vs. Standard distinction also reveals a product hypothesis — that developers will pay more for lower latency on complex tasks, which is a proxy for "AI coding time is productive time worth paying for."

Critical questions:

  • Does Composer 2 consume from the existing monthly request quota, or is it entirely separate — and if separate, does it replace or supplement the existing Composer model?
  • What's the average session cost for a typical Composer 2 workflow — without that, developers can't evaluate whether $7.50/M output tokens is cheap or expensive relative to the value delivered?
  • Is there a cap or alert mechanism so teams don't inadvertently rack up large bills on heavy Composer 2 usage?
  • What benchmark or eval does "strong results on challenging coding tasks" refer to — is this a controlled comparison or marketing language?

Action you could take today: If your team uses Cursor, pull the last 30 days of usage data and estimate what Composer 2 Fast pricing would cost at your team's current token consumption. Compare that against the value of the highest-stakes tasks (debugging complex issues, refactoring large codebases) — that ratio tells you whether Composer 2 is worth the premium or whether Standard tier is the right default.

Wispr Flow — Android Launch with Free Unlimited Dictation#

Source: https://play.google.com/store/apps/details?id=com.wispr.flowapp (via roadmap.wisprflow.ai changelog) Credibility: High (first-party product changelog entry)

What happened: Wispr Flow launched its Android app in early access with a significant introductory offer: free, unlimited dictation for all users for a limited time. The changelog explicitly frames this as "not just a 'mobile version' of Flow" — it's described as the same core dictation system that runs on desktop, now optimized for Android's input architecture. This is a notable platform expansion for a tool that has been primarily desktop-focused (macOS) and used by professionals at companies like OpenAI, Vercel, and Nvidia.

Key details:

  • Full Wispr Flow experience on Android — not a stripped-down mobile companion
  • Free unlimited dictation during the early access period — a deliberate acquisition strategy to build Android install base
  • "Optimized for Android's input architecture" — not a port, but a rebuild for the platform, which suggests thoughtful mobile-native design decisions
  • Wispr Flow has previously restructured team pricing (covered Feb 28) — the Android launch continues a pattern of growth-mode product decisions in 2026

Why it matters for PMs: Platform expansion from desktop to mobile is a meaningful strategic signal, not an incremental update. Wispr Flow is betting that voice-first AI input isn't just a desktop power-user behavior — it's a mobile-native workflow waiting to happen. The free unlimited dictation offer is a classic land-and-expand move: get Android users hooked during early access, then convert to paid Flow Pro. For PMs tracking voice input modalities, this is the first sign that Wispr Flow is competing for mobile use cases where typing is awkward (commuting, walking, quick async messages). The open question is whether mobile dictation accuracy and workflow integration (Android's fragmented app ecosystem vs. macOS's tight accessibility APIs) meets the quality bar that professional users expect.

Critical questions:

  • How does dictation accuracy on Android compare to macOS — is the underlying model the same, or has it been tuned differently for mobile?
  • What's the "limited time" window for free unlimited dictation, and what's the conversion strategy to paid plans after it ends?
  • Which Android apps does Flow integrate with natively — does it work across email, Slack, Notes, and third-party tools the way the desktop version does?
  • What's the target user for mobile Flow — existing desktop users who want mobile parity, or new users who primarily work on mobile?

Action you could take today: If your team uses Wispr Flow on desktop, share the Android early access with team members who frequently draft async messages on mobile (Slack, email, docs). Collect qualitative feedback on one specific use case: does voice dictation on Android feel as reliable as typing for professional communication? That signal will tell you whether mobile voice-first workflows are ready for your team or still early.

Quick Hits#

  • Dan Shipper (Every): "When your vibe-coded app goes viral and then goes down" — firsthand account of Proof's launch-day server crashes, with the key line: "If you can vibe code it, you can vibe fix it. You just might not be able to fix it quickly." Concrete warning about operational risk in AI-assisted development. (March 20, 2026): https://every.to/chain-of-thought/when-your-vibe-coded-app-goes-viral-and-then-goes-down
  • Ben Tossell (Ben's Bites): "What makes a good AGENTS.md?" — analysis of the emerging AGENTS.md convention for coding agents (equivalent to CLAUDE.md or Cursor rules), covering community examples and what distinguishes effective agent instructions from ineffective ones. (March 18–19, 2026): https://www.bensbites.com/p/what-makes-a-good-agentsmd
  • Mistral AI: Launched Forge — a system that lets enterprises build "frontier-grade AI models grounded in their proprietary knowledge." Positions as a custom model platform for enterprise, separate from API access to existing models. (March 17, 2026): https://mistral.ai/news/forge
  • Karri Saarinen (Linear): Published "The Malleable Software That Never Was" — argues that the long-held dream of infinitely customizable software is obsolete now that agents can handle complexity, and that opinionated purpose-built software paired with agents beats user-shapeable tools. (March 20, 2026): https://www.linkedin.com/pulse/malleable-software-never-karri-saarinen-me7gc
  • Character.AI: Launched Imagine Gallery — a new image-focused feature described as a gallery product, published March 18. Suggests Character.AI is expanding beyond text-based character interaction toward visual/generative content. (March 18, 2026): https://blog.character.ai/imagine-gallery/

This Week's Pattern#

Agents are moving from developer infrastructure to mainstream productivity products. This week: Notion ships Custom Agents for knowledge workers with a full security architecture; Google launches a full-stack vibe coding agent in AI Studio; LangChain rebrands Agent Builder as Fleet for enterprise teams; GitHub documents Squad for multi-agent repository coordination. The pattern is not "agents are coming" — it's "agents have shipped into products that non-developers actually use." The infrastructure questions (sandboxes, execution environments, orchestration) are being answered; the product questions (trust, workflow fit, error recovery) are now the frontier.

Reflection Prompt#

Notion launched Custom Agents with four simultaneous blog posts: the announcement, the security model, internal usage at Notion, and a customer case study. The security post explicitly documents permission scoping and isolation testing done before GA.

For your agent product or feature: When you ship an agent that takes autonomous actions on behalf of users, what's your equivalent of Notion's security post — the artifact that documents how you thought about permission boundaries, failure modes, and what happens when the agent acts on bad data? If you don't have one yet, that's the gap between an agent demo and an agent product.

Complete your reflection in /content/reflections/daily/2026/2026-03-21.md