Home
May 19, 2026
View All

Cursor Composer 2.5, Vercel Sandbox Agents, and How Engineers Actually Use Claude Code

·1 underrepresented voice

The Short Version#

Three interconnected signals today: Cursor shipped a meaningful agent upgrade, Vercel made it easier to run Claude agents in sandboxed environments, and Thariq Shihipar from Anthropic gave a rare look at how engineers actually use Claude Code day-to-day — HTML over markdown, micro-apps for specs, and a reframe of what the PM role even means when you have compute to allocate.

Cursor — Composer 2.5#

Source: https://cursor.com/changelog/composer-2-5 Credibility: High (first-party changelog)

What happened: Cursor shipped Composer 2.5, a substantial upgrade to its agentic coding mode. The changelog describes it as "a substantial improvement in intelligence and behavior over Composer 2" — better at sustained work on long-running tasks, more reliable at following complex instructions, and "more pleasant to collaborate with." That last phrase is doing real work: it's not a benchmark claim, it's a behavioral one.

Key capabilities:

  • Improved performance on long-running, multi-step tasks (the core agentic use case)
  • More consistent instruction-following across complex prompts
  • Framed explicitly as a collaboration improvement, not just a capability jump

Why it matters for PMs: Composer 2.5 is a direct response to the biggest friction point in AI coding agents: they're great for short bursts and fall apart on anything that requires sustained context and judgment over many steps. If this holds up in practice, it shifts what's actually possible to delegate — not just single-file edits but multi-file, multi-step work. For PMs who use Cursor themselves or manage teams that do, the question changes from "what can I use AI for" to "what should I still be doing myself."

Critical questions:

  • "More pleasant to collaborate with" is hard to verify from a changelog alone. What's the failure mode it's fixing — mid-task confusion, unnecessary clarification loops, context drift?
  • Does the improvement hold for non-code tasks (spec editing, documentation) or is it primarily a code generation win?
  • How does this compare to Windsurf's Cascade agent on the same long-running task types?

Action you could take today: If your team uses Cursor, pick one task you've previously found the agent abandoned or got wrong mid-way through, and run it again with Composer 2.5. Concrete regression test beats any benchmark.

Vercel — Run Claude Managed Agents with Vercel Sandbox#

Source: https://vercel.com/changelog/run-claude-managed-agents-with-vercel-sandbox Credibility: High (first-party changelog)

What happened: Vercel shipped a new capability letting you run Claude-managed agents directly inside Vercel Sandbox — their isolated execution environment for untrusted or agentic code. This lands alongside a Flat Rate CDN beta (fixed monthly fee replacing usage-based pricing) and a firewall change that makes firewall-mitigated traffic free. The agent sandbox is the one with real product implications.

Key capabilities:

  • Claude agents can execute inside Vercel's sandboxed runtime
  • Sandboxing provides isolation for agentic workloads — important when agents are running code or making external calls
  • Integrates with existing Vercel deployment and infrastructure

Why it matters for PMs: Running agents in production has a trust and safety problem: agents that execute code or interact with external services need containment. Most teams either skip that containment (risky) or build it themselves (slow). Vercel just removed that build-it-yourself step for teams already on their platform. This lowers the activation energy for shipping agentic features considerably. The Flat Rate CDN is also worth flagging separately — it solves a real budget predictability problem for teams whose AI features generate unpredictable traffic spikes.

Critical questions:

  • What are the actual constraints of the Vercel Sandbox environment? What can Claude agents do, and what's blocked?
  • Is this a competitive response to similar sandboxing from AWS Bedrock AgentCore or just a natural platform extension?
  • The Flat Rate CDN is in "limited beta for Pro teams" — what's the pricing, and does it actually pencil out vs. usage-based for variable AI traffic?

Action you could take today: If you're building or evaluating an agentic feature on Vercel, check whether Sandbox support changes your security/containment architecture. It might remove a sprint of infrastructure work.

Thariq Shihipar (Anthropic) via Lenny's Newsletter — How Anthropic Engineers Actually Build with Claude Code#

Source: https://www.lennysnewsletter.com/p/html-is-the-new-markdown-how-anthropic Credibility: High (first-party interview with Anthropic's Claude Code engineer, published May 18)

What happened: Thariq Shihipar, an engineer on Claude Code at Anthropic, sat down with Lenny's newsletter to explain how Anthropic's own engineers use Claude Code in practice. Two things stand out. First: HTML has replaced markdown as the output format of choice. Instead of having Claude generate a markdown spec document, engineers prompt it to generate a small interactive HTML app — a "micro-app" — for things like editing specifications, exploring design decisions, or visualizing system behavior. Second: Thariq uses the phrase "compute allocator" to describe what engineers are becoming. You're no longer writing code; you're deciding where to apply Claude's output.

Key patterns:

  • HTML over markdown: Markdown is static and linear. An HTML micro-app can have forms, conditionals, toggles — it becomes an interactive artifact, not a document. Engineers use these for spec editing and design reviews.
  • Living design systems: Instead of maintaining a separate design doc, engineers generate micro-apps that are the design system interaction — you click through the spec rather than read it.
  • "Compute allocator" framing: Your job shifts from writing to deciding: which tasks get Claude's attention, which outputs get accepted, where does human judgment matter.
  • Claude Code Show and Tell events: Teresa Torres separately mentioned she's running these events (next one May 26th), which suggests this pattern is spreading beyond Anthropic internally.

Why it matters for PMs: The compute allocator framing is the most PM-relevant insight here, and it's uncomfortable in the right way. PMs have always been in the business of deciding where to invest scarce attention. If AI dramatically reduces the cost of execution, the job becomes even more about prioritization and judgment, not less. The HTML micro-app pattern is also directly applicable to PM work: spec reviews, user flow explorations, and prioritization frameworks that would normally live in a static doc could be interactive artifacts instead. Teresa Torres running Claude Code Show and Tell sessions signals this is moving from early-adopter to practical workflow territory for product people.

Critical questions:

  • The HTML micro-app pattern assumes Claude can generate reliable interactive artifacts. How much prompt engineering is required to make that work consistently, and does it generalize outside of engineering contexts?
  • "Compute allocator" is a seductive reframe, but it elides a real question: what happens to craft when your job is allocating rather than making? What's lost?
  • How does this pattern change code review and knowledge transfer on teams? If specs are micro-apps generated by Claude, who owns them?

Action you could take today: Next time you're writing a static spec or design doc, try prompting Claude to generate an HTML page with interactive elements instead. Even a simple form with conditional logic shows you something a bullet-point doc can't.

Quick Hits#

The Thread#

The "how do engineers actually use this" question is becoming more important than the capability question. This week surfaced three concrete practitioner accounts: Thariq on HTML micro-apps replacing markdown specs, Teresa Torres running Claude Code Show and Tell sessions for PMs, and Composer 2.5 shipping explicitly around behavioral improvements rather than benchmark wins. The pattern: the interesting product signal has shifted from "what can AI do" to "how do people actually work with it, and what does that change about the job."

Sit With This#

Thariq Shihipar describes Anthropic engineers generating interactive HTML micro-apps instead of static markdown specs — artifacts you can click through rather than documents you read. The "compute allocator" framing is the deeper claim: your job becomes deciding where Claude's output goes, not producing output yourself.

For your product team: If spec-writing and documentation are tasks you or your team already delegate heavily to AI, what part of your PM work do you believe is genuinely irreplaceable by that logic — and how would you know if you were wrong?