Home
Aug 1, 2026
View All

Perplexity Goes Multiplayer, AI Token Budgets Arrive, and Stateless MCP

·1 underrepresented voice

The Short Version#

Three signals today that all point at the same underlying question: who controls AI resource consumption, and at what level of the stack? Perplexity just made Computer a multiplayer agentic OS. Atlassian is capping AI token spend per employee. And Simon Willison's stateless MCP work shows why the protocol's architecture choices matter more than most PMs realize.

Perplexity Computer — Projects Launch Makes It a Multiplayer Agentic OS#

Source: https://x.com/AravSrinivas (pinned post, August 1, 2026) Credibility: High (first-party announcement from CEO, Aravind Srinivas)

What happened: Aravind Srinivas announced Projects on Perplexity Computer today, framing it as turning Computer into "a multiplayer agentic operating system with persistent memory and file systems." This is now available to all users, not just a waitlist. The shift is meaningful: Perplexity Computer launched as a solo tool for computer-use tasks, and Projects makes it collaborative and stateful at the team level.

Key capabilities:

  • Persistent memory across sessions, scoped to a Project
  • Shared file systems accessible across team members
  • Multiplayer mode for agentic workflows (multiple users collaborating inside the same agent context)
  • Available to all users immediately

Why it matters for PMs: This is Perplexity making the same bet Wispr Flow made when it launched Team Dictionary: the path from "individual power tool" to "team infrastructure" is through shared state. A solo agent that can't remember context across sessions is a novelty. One that maintains files, memory, and workspace across a team starts to look like a new kind of operating environment. The PM question here is whether Perplexity can hold this position against ChatGPT's projects feature and Notion's agent integrations, both of which are fighting for the same "team AI workspace" category. Perplexity's advantage is that Computer already does real computer-use tasks — it's not just chat with a label on it.

Critical questions:

  • How is persistent memory scoped? If it's per-Project but not role-gated, there are real data exposure risks for enterprise teams.
  • Does "multiplayer" mean simultaneous collaboration, or just shared context that multiple people can access sequentially?
  • What's the differentiation from Notion's custom agents (which also allow agentic workflows with shared workspace context)?
  • Is this targeting consumers, prosumers, or actual enterprise teams? The answer determines whether the file system abstraction is enough or whether you need SSO, admin controls, and audit logs.

Action you could take today: Map out the "team AI workspace" competitive landscape across Perplexity Computer, Notion Custom Agents, and ChatGPT Projects. What capability does each own, and what's the switching cost? That's the space Perplexity just moved into — and it's getting crowded.

Dare Obasanjo on Atlassian's AI Token Budgets — The Tokenmaxxing Era Ends#

Source: https://mas.to/@carnage4life (July 30, 2026, referencing The Guardian article) Credibility: Medium-High (Dare Obasanjo's commentary on a reported policy, not a primary source from Atlassian)

What happened: Atlassian has implemented per-employee AI token budgets of $500-$2,000/month. Dare Obasanjo's read on this is sharp: he calls it the end of the "tokenmaxxing era" and frames it as a compensation differentiation signal — employees who use AI efficiently will be valued differently than those who burn tokens on poorly constructed prompts. The policy implies Atlassian is treating AI token spend as a measurable resource, not an overhead cost, and tying it to individual output expectations.

Key patterns:

  • $500-$2,000/month range suggests significant variance by role or function
  • Framing is cost control, but the implication is productivity accountability: if you have a budget, you're expected to show returns
  • "Tokenmaxxing" culture (prompting at scale without discipline) was previously invisible to management; budgets make it legible
  • This also creates a new compensation lever: higher token budgets as a perk or performance signal

Why it matters for PMs: This is a preview of how AI costs get operationalized inside large organizations. Right now, most teams treat AI API costs as a line item in infrastructure — invisible to individual contributors. Atlassian's move makes it personal. For PMs building products that embed AI, this raises a concrete question: if your users are inside companies with token budgets, does your product's AI usage count against their budget or the company's? That distinction will matter for enterprise procurement and usage design. It's also worth watching how this plays out culturally — token budgets are only as good as the feedback loop that tells employees whether their spending was worth it.

Critical questions:

  • Is the $500-$2,000 range role-differentiated (engineers vs. support vs. marketing) or performance-differentiated?
  • How does Atlassian plan to close the feedback loop between token spend and measurable output?
  • For SaaS products with AI features: are enterprise customers starting to ask for per-seat AI consumption data so they can manage against internal budgets like this?
  • Does this create a market for "token-efficient" AI features as a differentiated product positioning?

Action you could take today: Check whether your product or team's AI usage has a visible cost attribution. If it's buried in infrastructure spend, consider what it would look like to make it legible at the individual or team level — and what behaviors that would change.

Simon Willison — Stateless MCP Recaptures His Interest#

Source: https://simonwillison.net/2026/Jul/31/stateless-mcp/#atom-everything (July 31, 2026) Credibility: High (Simon Willison is one of the most rigorous practitioners covering MCP architecture)

What happened: Simon Willison published a post explaining why the new stateless variant of MCP has re-engaged him after earlier skepticism. He also shipped two related tools: mcp-explorer (for exploring MCP servers) and datasette-mcp (an MCP server for Datasette). The stateless MCP spec — released July 28 per Vercel's changelog (which already updated to support it) — changes the architecture so MCP servers don't need to maintain persistent connections. This resolves the main deployment headache with the original protocol.

Key technical details:

  • Original MCP required persistent server connections, which made hosting and scaling awkward
  • Stateless MCP lets servers handle individual requests without maintaining session state
  • This makes MCP servers much easier to deploy behind standard HTTP infrastructure (no WebSocket management, no sticky sessions)
  • Vercel's changelog confirms they updated Vercel MCP to support the 2026-07-28 spec the same day
  • Simon's mcp-explorer tool lets you inspect what any MCP server exposes — useful for evaluation and debugging
  • datasette-mcp is a concrete example: a read-only data exploration interface over any SQLite database, now accessible to any MCP-compatible client

Why it matters for PMs: MCP's adoption has been bottlenecked by the infrastructure complexity of the stateful protocol. If stateless MCP simplifies deployment enough that any team can stand up an MCP server without specialized infrastructure, the ecosystem could expand quickly. For PMs deciding whether to build MCP-compatible integrations, this is the signal that the protocol is stabilizing in a direction that's actually deployable. The Vercel same-day update is a strong adoption signal — it suggests the new spec was coordinated, not just proposed. Watch for other platforms updating in the next two weeks.

Critical questions:

  • Does stateless MCP sacrifice anything important? Specifically: how are long-running tool calls handled if there's no persistent session?
  • What's the upgrade path for existing MCP servers that rely on session state?
  • If MCP deployment gets easy enough, does it commoditize the "connect your tools to AI" space, or does the value shift to which tools you connect?
  • Simon is one of the most thoughtful critics of AI infrastructure hype — his re-engagement is meaningful. What specifically changed his mind?

Action you could take today: Read Simon's mcp-explorer output for one MCP server your team is evaluating or already uses. The gap between what a server advertises and what it actually does well is often where integration quality breaks down.

Quick Hits#

  • Cassie Kozyrkov: "When AI Makes You Aggressively Ignorant" — post on how AI tools can create confident blind spots rather than actual knowledge. Directly relevant for PMs thinking about how AI-assisted workflows change what users know vs. think they know. (August 1, 2026): https://kozyrkov.medium.com/when-ai-makes-you-aggressively-ignorant-32542e085863

  • Notion: AI Meeting Notes can now trigger Custom Agents — meeting summaries can now kick off automated agent workflows. This is a meaningful workflow integration, not just a feature add. (July 31, 2026): https://www.notion.so/releases/2026-07-31

  • Vercel AI Gateway: Now supports team and project spend budgets with alerts — ships the same day Atlassian's token budget story breaks. The tooling for AI cost governance is arriving at the infrastructure layer and the enterprise policy layer simultaneously. (July 31, 2026): https://vercel.com/changelog/ai-gateway-spend-budgets-and-alerts

  • Patrick Collison: New business formation on Stripe has roughly doubled year-over-year — the largest YoY jump in Stripe's history. Collison attributes it to AI lowering the cost of starting a company and predicts "many thousands of winners" rather than winner-take-all consolidation. (July 26, 2026): https://www.ycrootaccess.com/p/patrick-collison-what-if-you-succeed

  • OpenAI: Ten advances in mathematics and theoretical computer science — concrete results on open problems in geometry, cryptography, and complexity. Not a product launch, but relevant as a signal of where AI-assisted research is producing verifiable outputs rather than claims. (August 1, 2026): https://openai.com/index/ten-advances-in-mathematics

The Thread#

AI resource governance is becoming a real product category. This week: Atlassian caps per-employee token spend, Vercel ships spend budgets and alerts at the gateway level, LangSmith ships its LLM Gateway with runtime spend controls, and Perplexity adds persistent shared memory to Computer projects. These aren't coincidentally arriving together — the "AI is free to experiment with" phase of enterprise adoption is closing, and the "who owns what, at what cost" phase is opening. PMs building AI features for enterprise customers should expect procurement conversations to include token consumption data, budget attribution, and cost-per-outcome framing within the next two quarters.

Sit With This#

Atlassian is giving employees $500-$2,000/month in AI token budgets and — implicitly — expecting them to show returns on that spend. The budget makes AI usage legible in a way it wasn't before.

For your team or product: If your users had a visible monthly budget for AI feature usage, what behavior would change? And is the current invisible-cost model actually hiding whether your AI features are delivering value — for your users or for your own team?