Home
Jun 29, 2026
View All

Cursor Goes Mobile and Gusto Ships Without Docs

The Short Version#

Two patterns collide today: AI coding tools keep expanding their surface area (Cursor on iOS, Vercel adding voice agents), while a real-world case study from Gusto shows what a team actually looks like when they take these tools seriously enough to throw out the old PM playbook entirely.

Cursor — iOS Mobile App in Public Beta#

Source: https://cursor.com/changelog/ios-mobile-app Credibility: High (first-party changelog)

What happened: Cursor shipped a mobile app for iOS, now in public beta on all paid plans. You can launch and manage always-on cloud agents directly from your phone — pick a repo, start an agent, monitor it, and push it toward completion without touching a desktop.

Key capabilities:

  • Launch and manage cloud agents from mobile (same flow as desktop)
  • Start agents from a repo selection screen
  • Interact with running agents: add context, review suggestions, apply changes, request revisions
  • See all active agents across a project at a glance
  • Requires a paid plan (Pro or above) to access the beta

Why it matters for PMs: This is a meaningful surface area shift. Cursor started as a desktop code editor. Now it's becoming an agent management platform that travels with you. The strategic move here isn't "Cursor on your phone" — it's Cursor betting that running agents asynchronously, over hours, is how people will actually use AI coding in practice. If your agents can run unattended and you just check in to review and steer, a mobile interface makes sense. For PMs thinking about AI-assisted workflows on their own teams, this is worth watching: the premise is that meaningful coding work happens across a full day, not just in a focused editor session.

Critical questions:

  • Who actually wants to manage coding agents from their phone? Developers, or PMs and tech leads who want to monitor without coding?
  • Does async agent management require a new mental model from users, or does it map onto something familiar (like monitoring CI/CD pipelines)?
  • What happens to error recovery when the human is on a mobile screen? Are the interaction patterns sufficient?
  • Is the target user someone managing multiple agents across a team, or a solo developer who wants to check in while away from desk?

Action you could take today: If your team uses Cursor on paid plans, grab the iOS beta and run one real agent session end-to-end on mobile. Pay attention to where you reach for the desktop — that friction is the product signal.

Gusto — 5 People, 10 Weeks, No Docs, New Product Line#

Source: https://www.lennysnewsletter.com/p/no-figma-no-jira-no-docs-how-gusto Credibility: High (Lenny's Podcast, interview with Eddie Kim, Gusto CTO)

What happened: Gusto's CTO Eddie Kim detailed how a 5-person team shipped a full AI product line in 10 weeks using Claude Code — and explicitly dropped Figma, Jira, and written documentation from the process. The team ran a "perma-Zoom" instead: a persistent video call where people worked alongside each other in real time, with Claude Code handling the implementation layer.

Key patterns:

  • No Figma for design, no Jira for task tracking, no docs for async communication
  • Persistent video call replaced documentation and async coordination
  • Claude Code as the implementation engine — not just for writing code but as the connective tissue between decisions and execution
  • 5-person team compressed a typical product cycle by eliminating handoff overhead
  • The 10-week timeline included shipping a product that users actually interact with, not a prototype

Why it matters for PMs: This is the sharpest real-world example I've seen of what happens when a team decides to actually trust AI tooling rather than add it alongside existing processes. The instinct for most teams is to add Claude Code (or Copilot, or Cursor) on top of existing workflows. Gusto went the other way: they rebuilt the workflow around the tool. The tradeoff is real — perma-Zoom doesn't scale to large teams, and "no docs" creates institutional knowledge risk. But for a 5-person team with a clear goal and a time constraint, the calculus apparently worked. For PMs, the most useful question this raises is: which parts of your current process exist to coordinate humans, and which parts would you actually need if an AI agent could hold more of the execution context?

Critical questions:

  • What does onboarding look like when you have no docs and no Jira history? How does a new team member get up to speed?
  • Is the perma-Zoom model sustainable beyond 10 weeks, or is this a sprint pattern that accumulates debt?
  • How did they handle product decisions that needed stakeholder alignment outside the 5-person team?
  • What's the quality tradeoff? Shipping fast with Claude Code — what had to be reworked or didn't survive production?

Action you could take today: Pick one upcoming sprint where your team is small enough (3-5 people) and time-boxed enough to try dropping one layer of documentation in favor of more synchronous coordination. Track where the missing doc actually hurts you — that's your highest-leverage place to start with AI tooling.

Vercel — Realtime Voice Agents on AI Gateway#

Source: https://vercel.com/blog/realtime-voice-agents-on-ai-gateway and https://vercel.com/changelog/realtime-voice-speech-and-transcription-now-supported-on-ai-gateway Credibility: High (first-party blog and changelog)

What happened: Vercel's AI Gateway now supports realtime voice, speech, and transcription — meaning developers can build voice agent pipelines directly through the gateway without routing audio through separate infrastructure. xAI Grok audio models are also now available on the gateway as a provider option.

Key capabilities:

  • Realtime voice and speech-to-speech support routed through AI Gateway
  • Transcription services now available alongside LLM calls in a unified gateway
  • xAI Grok audio models available as a provider alongside existing options
  • Builds on Vercel's existing AI Gateway (unified routing, observability, rate limiting)
  • Designed to simplify the infrastructure for building voice agents

Why it matters for PMs: Voice agent infrastructure has been fragmented: you'd need separate services for STT, LLM, and TTS, then stitch them together. Vercel consolidating this under a single gateway with existing observability tooling lowers the bar for prototyping voice features significantly. If you're thinking about adding a voice interface to a web product, the "where do I even start" question just got simpler. The deeper signal is that Vercel keeps expanding AI Gateway from an LLM proxy into a full AI infrastructure layer — compute, voice, memory, sandboxes. That's a meaningful platform bet.

Critical questions:

  • What are the latency characteristics of routing realtime audio through a gateway versus direct provider connections? Voice agents are sensitive to even small delays.
  • Which specific use cases does this unlock that weren't viable before, versus making existing use cases slightly easier?
  • How does pricing work for audio tokens versus text tokens on the gateway? Is the cost model predictable enough for production voice features?

Action you could take today: If you've been evaluating voice AI features, check the Vercel AI Gateway docs for voice support — this is worth a prototype session if you're already in the Vercel ecosystem.

LangChain — Dynamic Subagents in Deep Agents#

Source: https://www.langchain.com/blog/introducing-dynamic-subagents-in-deep-agents Credibility: High (first-party blog)

What happened: LangChain shipped dynamic subagents for Deep Agents — a feature that lets AI agents orchestrate work at scale using code instead of tool calls. Rather than pre-defining a fixed set of subagent calls, an orchestrator agent can write code to spin up subagents dynamically, enabling fan-out at scale and reliable coverage for complex, multi-step pipelines.

Key technical details:

  • Orchestrator agents generate code that instantiates and coordinates subagents programmatically
  • Fan-out: one orchestrator can spawn many parallel subagents to cover all cases
  • Guarantees coverage across the full problem space (versus tool-call orchestration, which can miss cases)
  • Common orchestration patterns: scatter-gather, parallel processing, conditional branching
  • Live traces available to inspect how orchestration actually runs

Why it matters for PMs: The difference between "AI agent" and "reliable AI agent pipeline" comes down to whether the orchestration layer can guarantee coverage. Tool-call-based orchestration is brittle: the agent decides which tools to call, which means it can skip steps. Code-based orchestration is explicit and auditable. For PMs building agentic workflows — customer support, research pipelines, compliance review — this distinction matters a lot for what you can actually promise users in terms of reliability. It also connects directly to the open question of how much autonomy users will tolerate: a pipeline that provably covers all cases is easier to trust than one that might.

Critical questions:

  • What's the failure mode when the code the orchestrator generates has a bug? How do you catch that in production?
  • Does code-based orchestration make it harder to debug issues when something goes wrong mid-pipeline?
  • Is this pattern accessible to teams without strong engineering support, or does it require dedicated agent infrastructure expertise?

Action you could take today: If you have an agent use case with variable fan-out (e.g., checking multiple sources, processing multiple items), read LangChain's Deep Agents documentation on dynamic subagents and map it to your current architecture.

Quick Hits#

  • Dare Obasanjo: Commentary on Ford rehiring hundreds of senior engineers after AI replacement backfired — Ford now the most recalled automaker in the US with 51 recalls in H1 2026. Sharp skeptical signal on the limits of replacing experienced engineers with AI (2026-06-28): https://mas.to/@carnage4life/113270162749465864

  • Lenny Rachitsky / Gusto CTO Eddie Kim: Companion audio episode to the Gusto write-up above, covering GLM-5.2 review alongside the Claude Code case study. Good listen if you want the full context (2026-06-29): https://www.lennysnewsletter.com/p/how-i-ai-glm-52-review-and-how-gusto

  • Simon Willison: Flagged Ornith-1.0, a project about self-scaffolding LLMs for agentic coding — LLMs that generate their own scaffolding code rather than relying on pre-built frameworks. Relevant if you're tracking agent architecture patterns (2026-06-29): https://simonwillison.net/2026/Jun/29/ornith/#atom-everything

  • ElevenLabs: v1 TTS models (eleven_monolingual_v1, eleven_multilingual_v1) being removed July 9, 2026. If your product uses ElevenLabs voice, you need to migrate to eleven_multilingual_v2 or a current model before that date. Also: SynthID integration for detecting ElevenLabs-generated audio is now live (2026-06-25): https://elevenlabs.io/blog/synthid

  • Aravind Srinivas: Perplexity and Intel partnering to bring local models and hybrid inference to Intel Ultra Series 3 laptops — Perplexity positioning as the on-device AI interface alongside cloud. First real signal of Perplexity going after the local/hybrid inference use case (2026-06-29): https://x.com/AravSrinivas

The Thread#

AI development tooling is becoming infrastructure, not tooling. This week's pattern: Cursor (mobile app for agents), Vercel (voice on AI Gateway), and LangChain (code-based orchestration) are all making the same move — expanding from "tool that helps you do X" to "platform that runs your AI workflows end-to-end." The Gusto case study is the human side of that same pattern: when the tooling is capable enough, you don't add it to your process, you rebuild the process around it. The question for every PM right now is the same one Gusto answered for themselves: what am I doing to coordinate humans that an AI agent could hold instead?

Sit With This#

Gusto's team shipped a full AI product line in 10 weeks with 5 people by dropping Figma, Jira, and documentation entirely, replacing async handoffs with a persistent video call and Claude Code.

For your team: If you had to cut one coordination layer — docs, tickets, or design files — to move twice as fast on your current initiative, which one would you cut first? And what would you need to trust the AI tooling enough to actually do it?