Muse review: The personal AI agent that gets consumer UX right
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
Optimizely—Your AI agent orchestration platform for marketing and digital teams
OpenArt—An all-in-one AI creation platform for images, videos, music, audio, and more
In this solo episode, Claire tests Meta’s Muse, a personal AI agent designed for everyday users. She walks through how it manages her family calendar, creates a polished morning newsletter, tracks personal goals, and asks for permission before acting on sensitive information. She also explores the thoughtful details that make Muse feel trustworthy, where its browser use still falls short, and what its approach could mean for the future of consumer AI.
Biggest takeaways:
Meta’s consumer design expertise makes Muse feel different from other agents on the market. The teams behind Facebook and Instagram understand how to build trust with nontechnical users, and that experience shows up throughout Muse. It never exposed Claire to a terminal, surfaced a confusing error, or requested permission at an awkward moment. She tested onboarding, calendar management, email ingestion, goal setting, and browser-based purchasing in one session without encountering a single moment of friction that broke the experience.
Muse created the best version of Claire’s family morning brief on its first attempt. Claire has tested generating family newsletters with OpenClaw and Codex, but Muse produced a more beautiful and thoughtfully structured PDF without requiring any follow-up prompting. It even added a “Talk at the Table” section with kid-friendly questions about the news and proactively identified scheduling conflicts. That is the standard personal agents should aim for: not simply completing the request but noticing useful things the user did not think to ask about.
How an agent requests permission matters as much as what it can do once permission is granted. Muse does not default to unlimited access or interrupt the user with constant approval requests. It asks at the moment access becomes relevant, confirms what it found, and checks again before taking action. After reading Claire’s email to build a family profile, for example, Muse summarized what it had learned and asked whether it could use that information. Anyone building an agent that handles personal data should pay attention to that interaction pattern.
Muse’s activity feed provides the kind of agent transparency Claire wants everywhere. Every task includes a detailed record of the tool calls, scripts, searches, and individual steps Muse completed. Most consumers will probably ignore it, but developers and power users will find it invaluable. Claire immediately wanted the same interface in Codex and Claude Code. Seeing exactly how an agent arrived at an outcome builds a different level of trust than receiving a polished result with no visibility into the process.
Personal goals require a different product model than tasks or chat. Helping someone gently sleep-train an eight-and-a-half-month-old while room-sharing in a small house is not a request that can be completed in one sitting. Muse gathered context across multiple turns, established a reminder, and treated the situation as an ongoing goal whose progress should be tracked. Its tone was warm and specific without becoming overly agreeable. Claire said she felt “zero percent annoyed” with it, which she described as a miracle.
Browser use remains the weak link for consumer agents. Muse struggled to shop for New Balance 9060s in a specific colorway, returning incorrect results and failing to complete the purchase. It performed better while buying tickets for The Odyssey, but initially opened the wrong movie. These failures are not unique to Muse. They are a reminder that browser-based shopping remains unreliable across the entire agent category.
Muse’s animated avatar shows what it means to design specifically for AI. Instead of displaying a generic spinner while it works, Muse generates an animation in which its avatar picks up a tiny laptop when completing tasks and an orb when creating media. Nobody asked for this feature, but it makes the agent’s state immediately understandable. For Claire, this is what operating at the top of the craft looks like. It is not just polished corners and clean icons. It is using capabilities like image generation and animation to solve experience problems that could not have been solved before.
Muse is introducing a new vocabulary for consumer AI. It avoids technical terms such as crons, tools, artifacts, plugins, and connectors. Instead, it organizes the product around a feed, ideas, goals, and a library. That language is a deliberate design decision. The words a product uses determine who feels comfortable using it, and Muse is clearly designed for a much broader audience than today’s coding and productivity agents are.
Blog and detailed workflow walkthroughs from this episode:
Meta Muse Review: A Personal AI Agent for Family Life: https://www.chatprd.ai/how-i-ai/meta-muse-review-personal-ai-agent
↳ How to Create a Printable Family Newsletter with Meta Muse: https://www.chatprd.ai/how-i-ai/workflows/meta-muse-printable-family-newsletter
↳ How to Set Personal Goals and Reminders with Meta Muse: https://www.chatprd.ai/how-i-ai/workflows/meta-muse-personal-goals-reminders
↳ How to Find and Buy Movie Tickets with Meta Muse: https://www.chatprd.ai/how-i-ai/workflows/meta-muse-movie-tickets
How Warp ships 2,000 PRs a month with AI factories | Zach Lloyd (CEO, Warp)
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
DX—Engineering intelligence for the AI era
OpenArt—An all-in-one AI creation platform for images, videos, music, audio, and more
Zach Lloyd is the co-founder and CEO of Warp, an agentic development platform for software teams. In this episode, he demonstrates how Warp Factories turns a request in Slack into a tracked, tested, and merged pull request, with the first PR ready in as little as 35 minutes. He also explains how the team measures human involvement, scores every agent run, replays real tasks to compare models, and uses failed runs to continuously improve the factory itself.
Biggest takeaways:
A software factory is not just a faster coding agent. It is the entire development lifecycle running in the cloud. Warp’s factory, Wilson, can take a request from Slack through triage, Linear ticket creation, implementation, pull-request creation, computer-use QA, and merge without a human touching the keyboard. The difference between a factory and a coding agent is the difference between an assembly line and a single machinist. One completes a task. The other coordinates the entire production process.
“Human interactions per pull request” may be the clearest measure of factory efficiency. Every follow-up prompt in Slack, clarification on a Linear ticket, and correction during code review represents additional human effort. Zach believes that the more steering an agent requires, the lower the factory’s true throughput becomes. It does not matter how quickly the agent writes code if a human remains the rate limiter, yet most teams are not measuring this.
Humans in code review are now the biggest bottleneck in Warp’s engineering process. Wilson takes an average of 35 minutes to move from kickoff to pull request. The first human review arrives 3.5 hours later. Warp has already removed one source of delay by allowing the person who prompted the agent to review its work instead of requiring a separate reviewer. The company has not eliminated human review entirely, but Zach expects that accumulated trust will eventually make it possible for selected changes to merge without human review.
A company’s own past work provides a better model benchmark than a public leaderboard. Benchmarks such as SWE-bench and Terminal-Bench measure performance on generic tasks. Warp instead replays real factory jobs using different model configurations, evaluates them with the same judging system used in production, and maps the cost-quality tradeoff on a Pareto chart. For Warp’s workload, model selection is the largest cost lever by far, with context management a distant second.
Software factories can improve themselves when their configuration lives in code. Warp’s improvement loop collects failed runs until there is enough evidence to identify a pattern, typically around 20 to 25 examples. An observer agent analyzes what went wrong and proposes changes to the factory’s agent definitions. Because those definitions are stored as code, the same agents that complete the work can also update the instructions governing how they operate.
The scoring layer is what separates a reliable factory from a vibe-coded pipeline. Warp uses an AI judge to score every agent run across multiple dimensions, including whether it created redundant tests. Agents often write tests that merely confirm current behavior instead of protecting against future regressions. The judging model is intentionally mid-tier so evaluation costs do not erase the factory’s savings, and Warp can adjust the sampling rate depending on how much oversight a workflow needs.
AI becomes far more valuable outside engineering when it connects to real business data. During the conversation, Zach ran three tasks in parallel: redesigning a slide through Figma, extracting the 10 most common sales questions from four weeks of Granola transcripts, and creating a cold-lead re-engagement list in Google Sheets. The redesigned slide was useful, but the Granola analysis was more consequential. It identified “buy versus build” as the top customer question and backed that conclusion with evidence, giving Warp a clear signal about what its website and sales deck should address next.
Working with agents in public can spread expertise without a formal training program. Every Wilson task runs in a shared Slack channel, allowing junior employees to watch experienced users prompt, steer, and correct the system in real time. Zach sees a power-law effect in agent usage: a small number of people become dramatically better than everyone else. Making their workflows visible allows those skills to spread quietly across the organization.
Building faster does not help if the team builds the wrong thing. Warp still relies on user interviews, design sessions, and direct observation of how people use the product. The factory dramatically accelerates implementation, but it does not decide what deserves to be built or why. Those decisions still depend on human taste, customer understanding, and product judgment. The build phase may be changing quickly, but the work that comes before it remains remarkably traditional.
Blog and detailed workflow walkthroughs from this episode:
Inside Warp’s Software Factory: https://www.chatprd.ai/how-i-ai/inside-warps-software-factory
↳ How to Build an Automated Software Factory: https://www.chatprd.ai/how-i-ai/workflows/how-to-build-an-automated-software-factory
↳ How to Build an AI-Powered CEO Toolkit: https://www.chatprd.ai/how-i-ai/workflows/how-to-build-an-ai-powered-ceo-toolkit
↳ How to Measure and Self-Improve Your AI Software Development Factory: https://www.chatprd.ai/how-i-ai/workflows/how-to-measure-and-self-improve-your-ai-software-development-factory
If you’re enjoying these episodes, reply and let me know what you’d love to learn more about: AI workflows, hiring, growth, product strategy—anything.
Catch you next week,
Lenny
P.S. Want every new episode delivered the moment it drops? Hit “Follow” on your favorite podcast app.
Listened twice. I’m building a consumer agent, and this was pure gold. It’s already shaping how we think about the product. One small correction: you mention that it avoids technical terms, but I’ve noticed “artifacts” and “connectors” in the experience. That aside, such a useful review. Thank you for putting it together.