Claire Vo spends a lot of her time testing AI agents against each other. After a few hours with Muse, Vo said its activity feed was a feature she wished Claude Code and Codex had, and that its permission system felt different from other agents she’d used.
What she said
In her review of Muse for Lenny’s Newsletter, Vo describes putting the agent “through a real first-pass test: onboarding, calendar management, goal setting, a one-shot family morning newsletter, browser-based shopping, and the animated avatar that honestly surprised me.” A one-shot result is produced from one initial request, without iterative prompts. That alone does not establish its accuracy.
Her verdict: Muse is “the best-designed personal agent I’ve tested.” Two specifics stand out. She says Muse has an activity feed that shows “task lineage,” meaning you can see the trail of steps that led from your request to the agent’s result, not just the final answer. She then says this is a feature she “immediately wished Codex and Claude Code had.” She also describes Muse’s permission system, which governs what the agent can do on its own versus what it has to ask you about first, as “different from every other agent I’ve used.”
Who she is
Vo is the chief product officer at LaunchDarkly, a company that builds tools for controlling which users see which features in software. She has held the top product role at three companies (LaunchDarkly, Color Health, and Optimizely), and she also founded ChatPRD, a tool product managers use to write product specs with AI help. She hosts the podcast “How I AI,” where testing and comparing AI agents is a recurring format.
What she gets right, and where it’s incomplete
Vo’s two specifics are real product requirements, not cosmetic preferences. An agent that books tickets, sends emails, and spends money on your behalf needs a permission system you can actually understand, and a visible record of what it did and why. Those are the same concerns builders have been raising about coding agents that act with less oversight than people expect, including Cursor’s new Projects feature, which Cursor says can direct thousands of helper agents while maintaining context over months of work.
Her written summary highlights permission design and the activity feed; the linked episode includes dedicated walkthroughs of both. Before adopting a similar design, inspect which actions need approval and how the feed records them. A favorable review is not a security audit of those controls.
The categories being compared also differ. Vo tested Muse on personal tasks: a family newsletter, calendar management, shopping. She didn’t test it against Claude Code or Codex on writing or reviewing code, which is what those two agents are built for. Her claim is specifically about permission design and visibility, not about which agent is better at engineering work.
Why it’s notable
Meta launched Muse, its first consumer AI agent, on September 8. It’s a crowded field: Google, OpenAI, and a wave of startups are all shipping personal agents that act on a user’s behalf with varying degrees of oversight. Meta’s own developer-facing model, Muse Spark, already opened a limited preview to outside developers in July. Vo’s comparison lands at a moment when the agent tools builders use every day, including coding agents, are still working out how much an agent should do without asking first, and how clearly it should show its work.
What it means for builders
If you’re building or evaluating an agent product, define approval points and a visible activity trail before granting access to accounts or purchases. Vo’s review highlights how clearly users can understand those controls; their enforcement still needs testing.
End of article