Cursor Goes Parallel, OpenAI Adds Ads, and Perplexity Launches a Computer

Cursor — PR Review, Parallel Agents, and Split PRs#

Source: https://cursor.com/changelog/05-07-26 Credibility: High (first-party changelog)

What happened: Cursor's May 7 release (version 3.3) ships three things that compound: a new PR review experience, parallel agent execution on plans, and quick-action pills for common workflows. The parallel agents piece is the architectural shift — instead of working through a plan sequentially, Cursor can now spin up multiple agents executing different parts simultaneously. The PR review experience is available directly inside Cursor 3, and Split PRs is a new quick-action that lets you break large changes into smaller reviewable chunks.

Key capabilities:

• PR review in-editor: Review pull requests without leaving the coding environment — the context stays live
• Parallel plan execution: Plans that previously ran step-by-step can now execute sub-tasks concurrently, reducing time-to-completion on complex multi-file changes
• Split PRs: Automatic or assisted decomposition of large diffs into smaller PRs — directly addresses a common pain point for teams using AI to generate large chunks of code
• Quick-action pills: Surface common workflows (like "split PR") at the moment they're relevant, not buried in a menu

Why it matters for PMs: The parallel agents feature is the most significant architectural shift here. Up until now, even AI-assisted coding was mostly linear — the AI did one thing, then the next. Parallel execution is the same leap that made cloud computing feel fundamentally different from a single server. For PMs using Cursor or evaluating AI coding tools, this changes the equation on how long complex tasks actually take. The PR review and Split PRs features address a real adoption blocker: AI-generated code tends to produce large, hard-to-review diffs, which slows down the rest of the team even if the individual developer is faster. Cursor is now trying to solve the team friction, not just the individual friction.

Critical questions:

• How does parallel execution handle conflicts when two agents modify the same file? Is there a merge/resolution layer?
• Does the PR review experience have enough context about team conventions and standards, or is it generic code review?
• Split PRs is a useful feature — but does it understand logical separation versus arbitrary splitting? Bad splits create more review burden, not less.
• Does parallel execution increase token costs proportionally? For teams paying usage-based pricing, this could matter.

Action you could take today: If your team uses Cursor, pull up the changelog and test the Split PRs quick-action on the next large AI-generated diff. If the split quality is good, it's worth adding to your team's AI-assisted dev workflow documentation.

OpenAI — Testing Ads in ChatGPT#

Source: https://openai.com/index/testing-ads-in-chatgpt Credibility: High (official OpenAI announcement)

What happened: OpenAI announced it's beginning to test ads in ChatGPT. The stated rationale is supporting free access — ads fund the free tier the same way they do in search and social. According to the announcement, ads will be clearly labeled, answers will remain independent from ad relationships, there will be strong privacy protections, and users will have control. This is the first time OpenAI has formally moved toward advertising as a revenue model.

Key capabilities / constraints:

• Ads are clearly labeled and separated from answers
• OpenAI commits that answer quality/content is not influenced by advertiser relationships
• Privacy protections are mentioned but specifics aren't in the excerpt
• User control is mentioned — unclear what "control" means in practice (opt-out? ad preference settings?)

Why it matters for PMs: This is a business model decision with product implications that ripple outward. For PMs at companies that compete with ChatGPT on free-tier positioning, the calculus just changed — ads mean OpenAI can sustain a free tier indefinitely, which raises the floor for what "free AI" means for users. For PMs building on OpenAI's API, the question is whether ad-supported ChatGPT starts to feel different from API-based products — and whether users notice or care. The deeper question is trust: ads in search feel normal because search has always had ads. Ads in a conversational AI feel different because the medium is more intimate and the boundary between "answer" and "sponsored answer" is harder to perceive. If users start wondering whether ChatGPT's recommendations are paid placements, that's an opening for competitors who can credibly claim they aren't.

Critical questions:

• What does "answers remain independent from ad relationships" mean in practice, and how is it enforced? Who audits this?
• What user controls exist? Can free users opt out of ads entirely? Can they pay a small amount to remove ads?
• Does advertising change what topics ChatGPT is willing to engage with, even indirectly (e.g., avoiding content that advertisers wouldn't want adjacent to)?
• How does this affect enterprise and API customers who don't see ads — is there a differentiation angle there?

Action you could take today: If you're building a product that competes with or complements ChatGPT, draft a one-paragraph positioning statement for your team on how your product handles data and monetization differently. This is going to become a decision criterion for some users.

Perplexity — Launching "Perplexity Computer"#

Source: https://x.com/AravSrinivas/status/1851700699756925059 Credibility: Medium (CEO announcement on X, no detailed product page in collected data)

What happened: Aravind Srinivas announced the launch of Perplexity Computer, described as a general-purpose AI digital worker that "unifies every current capability of AI into a single system." The announcement is accompanied by a deprecation of the legacy Perplexity Mac App in favor of a new one featuring "Personal Computer." Based on the framing, this appears to be a move toward ambient/agentic computing — an AI that operates as a persistent presence on your machine rather than a tool you invoke for specific queries.

Key capabilities (based on announcement framing):

• Positions as a "digital worker" — implying task execution, not just information retrieval
• "Unifies every current capability" — suggests integration of search, answer engine, and likely agent/action capabilities
• New Mac App ships alongside the announcement, replacing the legacy app
• "Personal Computer" branding suggests ambient, always-available presence

Why it matters for PMs: Perplexity started as an AI search engine and has been steadily expanding the scope of what it does. "Perplexity Computer" is a significant positioning move — it's no longer competing with Google or ChatGPT on search/chat, it's competing with the operating system layer for how you interact with your computer. That's a much larger ambition and a much harder moat to build. For PMs, this is worth watching because: (1) it tests whether users want an AI at the OS layer versus a purpose-built tool, and (2) it signals that consumer AI products are racing toward ambient/always-on as the next product paradigm, not just better chat. The Mac App deprecation is also a real signal — they're willing to break backward compatibility to move fast toward the new vision.

Critical questions:

• What does "digital worker" actually mean in practice? What tasks can it complete autonomously versus with user confirmation?
• How does Perplexity Computer handle the trust/autonomy problem — what can it do on your machine without asking?
• Is this genuinely agentic (takes actions) or is it a rebranded search + chat experience?
• How does it differentiate from Apple Intelligence, which is also trying to operate at the OS layer?

Action you could take today: If Perplexity is a tracked competitor or adjacent product for your team, flag this announcement in your competitive landscape doc and set a calendar reminder to check back in 30 days — by then you'll have more concrete product details and early user reactions to evaluate.

GitHub — Improving Token Efficiency in Agentic Workflows#

Source: https://github.blog/ai-and-ml/github-copilot/improving-token-efficiency-in-github-agentic-workflows/ Credibility: High (official GitHub Engineering blog)

What happened: GitHub published a detailed post on how they're improving token efficiency in agentic GitHub Copilot workflows. The focus is on reducing token consumption in multi-step agent tasks — specifically addressing the problem that agentic workflows compound token usage at each step, which creates both cost and latency problems at scale.

Why it matters for PMs: Token efficiency is quickly becoming a real product constraint as teams move from one-shot AI interactions to multi-step agentic workflows. The cost of an agent that takes 20 steps is not the cost of 20 queries — it's often much higher because each step carries growing context. GitHub is being transparent about this problem, which is unusual and valuable. For PMs building agentic features or evaluating AI coding tools for their teams, this is the operational detail that gets glossed over in demos but bites you in production. Understanding the token efficiency patterns GitHub is solving also gives you a framework for evaluating other agentic tools.

Critical questions:

• What specific techniques are reducing token usage — context pruning, summarization, smarter retrieval?
• Does improving token efficiency come with any accuracy or reliability tradeoffs?
• How does this affect the pricing model for GitHub Copilot as it moves toward usage-based billing?

Action you could take today: Read the full post (linked above) and use it as a framework for evaluating token efficiency in any agentic tool your team is evaluating — ask vendors specifically how they handle context growth in multi-step tasks.