Home
Aug 5, 2026
View All

Perplexity Goes Multiplayer, Lenny Builds a Bot

The Short Version#

Two signals worth sitting with today: Perplexity just turned its Computer product into a team-native agentic OS, and Lenny Rachitsky built a working PR review bot in 30 minutes using Vercel Eve and Codex — and published the whole thing. Meanwhile, AWS shipped web search natively into Bedrock, which quietly closes a gap that's been pushing enterprise teams toward third-party grounding tools.

Perplexity — Projects Turns Computer Into a Multiplayer Agent OS#

Source: https://x.com/AravSrinivas Credibility: High (announcement from CEO Aravind Srinivas)

What happened: Perplexity launched Projects on Perplexity Computer, turning what was previously a single-user agentic environment into a shared workspace. Projects adds persistent memory, files, and sessions that are scoped across hubs and users. The framing from Aravind Srinivas is explicit: this is "a multiplayer agentic operating system for work." That's not incremental — it's a category repositioning.

Key capabilities:

  • Persistent memory scoped to a project, not a single session
  • Files and context shared across team members (hubs)
  • Sessions that persist across users, enabling async agentic work
  • Multi-user access to the same agent context

Why it matters for PMs: Individual AI tools face a structural ceiling: they're easy to adopt personally but hard to scale into team workflows because context doesn't carry. What makes tools sticky at the org level is shared infrastructure — shared memory, shared files, shared state. Perplexity just built that. This is the same move Wispr Flow made with Team Dictionary and Team Snippets, the same move Notion made when it went from personal workspace to team wiki. The playbook is: get individual power users, then give them a reason to pull their whole team in. Worth watching how Perplexity prices this — shared sessions and persistent memory are real infrastructure costs, and the answer there will tell you a lot about how confident they are in enterprise conversion.

Critical questions:

  • How does Perplexity handle access control and permissioning within a project? Enterprise teams will need granular controls before they trust shared agent memory.
  • Does persistent memory across users create data leakage risks between team members or hubs? What's the privacy model?
  • Is this a paid-tier feature only, or is it available on existing plans? The pricing structure will determine adoption ceiling.
  • How does this compete with OpenAI's team/enterprise offering, which also has persistent context? What's the differentiated angle?

Action you could take today: If you're building or evaluating team AI tools, map out where your product's context currently lives and whether it's user-scoped or team-scoped. The gap between those two is exactly the gap Perplexity just exploited.

Lenny Rachitsky — Built a PR Review Bot in 30 Minutes with Vercel Eve#

Source: https://www.lennysnewsletter.com/p/build-an-ai-code-review-bot-in-30 Credibility: High (first-person account from a practitioner PM with a large audience; published today)

What happened: Lenny built "Merge Mommy" — a PR review bot that scores risk, auto-approves low-risk PRs, and pings him in Slack for anything that needs a human. He used Vercel Eve agents and Codex, and documented the whole workflow. The headline claim: 30 minutes to a working bot. This is less a product announcement and more a real-world signal about what the current generation of agent tooling can do in a PM's hands.

Key technical details:

  • Built on Vercel Eve agents (Vercel's agent runtime, which also got browser access this week — see Quick Hits)
  • Used Codex for code generation within the workflow
  • Bot scores PR risk, auto-approves easy ones, escalates the rest to Slack
  • End-to-end from zero to working bot: ~30 minutes

Why it matters for PMs: This is the vibe coding pattern matured into real workflow automation. It's not "I made a demo" — it's a bot that runs on every PR and changes how a real person works. For PMs specifically, this kind of workflow automation is increasingly within reach without an engineering sprint. The 30-minute claim is worth taking seriously: Vercel Eve + Codex is now a tool combination that produces working agents quickly enough that PMs can prototype their own internal tooling. That changes what "PM productivity" means.

Critical questions:

  • How does "Merge Mommy" handle false negatives — auto-approving a risky PR that scored low? What's the error recovery path?
  • Is 30 minutes reproducible for a PM without Lenny's level of technical familiarity? What's the actual floor?
  • Does this hold up at scale — say, 50 PRs a day — or does latency and cost become a factor with Eve agents?
  • What happens when the bot gets a PR it can't categorize? Is there a fallback or does it just silently pass?

Action you could take today: If you have a repetitive review or triage task that costs you 10+ minutes a day, sketch out the decision logic as a simple if/then flow. Then look at whether Vercel Eve or a similar agent tool could execute it. The build time is shorter than you think.

AWS Bedrock — Web Search Is Now a Native Built-In Tool#

Source: https://aws.amazon.com/blogs/machine-learning/introducing-web-search-on-amazon-bedrock-for-foundation-model-grounding/ Credibility: High (official AWS Machine Learning blog, general availability announcement)

What happened: AWS launched Web Search on Amazon Bedrock as generally available — a server-side built-in tool that grounds model responses in current web knowledge, running entirely within AWS. It supports OpenAI GPT-5.4, GPT-5.5, and GPT-5.6 (Sol/Terra/Luna variants). The key detail: no third-party vendors to onboard, no external APIs, no data leaving AWS infrastructure.

Key capabilities:

  • Native server-side web search built into Bedrock — no external grounding service required
  • Works with OpenAI GPT models hosted on Bedrock (GPT-5.4, 5.5, 5.6 variants)
  • Grounding stays entirely within AWS, which matters for compliance and data residency
  • Removes the integration overhead of tools like Bing Search API, Tavily, or Brave that teams typically bolt on

Why it matters for PMs: Enterprise AI products that need current information — not just model knowledge — have historically required a third-party search grounding integration. That's an extra vendor, extra cost, extra data flow to manage, and extra compliance surface. AWS just eliminated all of that for teams already on Bedrock. This shifts the build vs. buy calculus: if you're on AWS, the answer for web grounding just became "use the native tool" instead of "evaluate and integrate an external service." It's also a signal that Bedrock is accelerating toward being a complete agent infrastructure stack, not just a model router.

Critical questions:

  • What are the rate limits and pricing for Web Search calls on Bedrock? Native doesn't mean free, and at scale this could be meaningful cost.
  • How does the quality of results compare to established grounding tools like Tavily or Bing? "Runs within AWS" is a compliance win but may not be a quality win.
  • Is this available on all Bedrock regions, or just select ones? Data residency is a key reason enterprises would use this, so regional availability matters.
  • Does this work with non-OpenAI models on Bedrock, or is it currently GPT-model-only?

Action you could take today: If your team is using a third-party search grounding service with Bedrock, check whether the new native Web Search tool covers your use case. Even a quick cost-and-compliance comparison could justify a migration.

Quick Hits#

  • Lenny Rachitsky: "Build an AI code review bot in 30 minutes with Vercel Eve" — detailed walkthrough of building Merge Mommy, a PR risk-scoring bot that auto-approves and Slack-pings (2026-08-05): https://www.lennysnewsletter.com/p/build-an-ai-code-review-bot-in-30

  • Vercel: Eve agents can now browse the web — "Give your eve agent a browser" is live in changelog, enabling agent workflows that require real-time information retrieval (2026-08-04): https://vercel.com/changelog/give-your-eve-agent-a-browser

  • Simon Willison: Released LLM 0.32 with support for reasoning traces, OpenAI Responses API, server-side tools, and smarter logging — useful for anyone evaluating or building on LLM CLI tooling (2026-08-04): https://simonwillison.net/2026/Aug/4/llm/#atom-everything

  • Pieter Levels: "Claude's AI output is so verbose and jargon dense that I have to look up every word" — a practitioner critique worth reading. When power users start complaining about output quality UX, that's a real signal about the gap between capability and usability (2026-08-04): https://levels.io/claude-ai-output-verbose-jargon-dense

  • Character.AI: Launched "(c.ai) series" — studio-made microdramas, a new content format. This is Character.AI moving toward original content production, which is a meaningful strategic shift from pure UGC toward platform-curated experiences (2026-08-04): https://blog.character.ai/cai-series/

The Thread#

The individual-to-team transition is the defining AI product move right now. Perplexity adding shared projects and persistent memory to Computer, Wispr Flow's Team Dictionary, Notion's agent integration, Cursor's Slack improvements — every major individual AI tool is making the same bet: the real retention and revenue is at the team layer, not the personal layer. The pattern is consistent enough now that if your AI product is personal-only, it's worth asking whether you're leaving the stickiest part of the market on the table.

Sit With This#

Lenny built a working PR review bot in 30 minutes using Vercel Eve and Codex. The bot scores risk, auto-approves the easy ones, and escalates the rest to Slack. It changes how he works — every day, on every PR.

For your product: What repetitive decision or triage task on your team currently takes a human 10-30 minutes per day and follows a consistent-enough logic that it could be expressed as rules? And what's actually stopping you from automating it this week?