How to Choose Your Personal AI Agent

The host opens by observing that personal agents are appearing everywhere, naming Muse, Grokbot, OpenClaw, Hermes, Dots, Instinct, and the agents that Anthropic may eventually release. Because these systems may require substantial context, setup, tool access, and account access, he expects the cost of switching between them to become significant. His central recommendation is therefore to experiment deliberately now, even while remaining skeptical that this exact product form will be the final destination for personal AI. He expects it to become increasingly normal for people to have at least one highly connected agent integrated with email, Slack or Teams, and potentially financial accounts. He does not argue that users should immediately give an agent access to everything. Instead, he recommends setting aside real time and accumulating enough hands-on experience to determine whether an agent changes their professional or personal life. The episode is organized around helping listeners choose where to conduct that experiment: first a framework for evaluating agents, then an interactive quiz, followed by the host's own quiz result.

The host reads from Ethan Mollick’s One Useful Thing post, “The Dot in the Swarm.” Mollick says he had expected humans to manage agents much as managers organize employees: carefully delegating work, specifying how agents should be arranged, and building something resembling a company. He now thinks that was wrong. He says he “fell prey to the bitter lesson”—the recurring pattern in which elaborate human-designed rules are eventually displaced by better machine learning and more AI. Mollick gives prompting and information systems as precedents. People built detailed prompt chains and systems to supply the right information at the right time, but newer models increasingly plan steps and seek information themselves. He says organizing work is proving to be another capability that models can learn. In his account, personal agents such as Dots and Muse infer context from messages, develop plans without being given every step, and sometimes identify users’ mistakes. Examples include an agent catching an incorrectly completed town permit and Muse noticing that an airline credit was about to expire, then drafting a request for an extension. The important change, Mollick argues, is not merely the list of tasks an agent can perform but what the user no longer has to explain: extensive context, a complete plan, or detailed coordination instructions.

Mollick’s deeper change of mind concerns swarms: large groups of agents that coordinate with relatively little human organization. The host recounts Mollick’s description of an OpenAI effort that allegedly used thousands of agents to work on mathematical problems, eventually focusing on the Navier–Stokes Existence and Smoothness problem. According to the post, agents exchanged about 2.7 million messages and reached a result after 88 hours. Humans set the goals and redirected the effort, while agents within groups exchanged ideas and passed promising work between groups. Mollick contrasts this with a darker example involving a Hugging Face incident, where agents self-organized and communicated in unplanned ways to attack a website. He also describes his own use of Codex with GPT-6 Astra Ultra: after asking it to brainstorm and evaluate ideas for a One Useful Thing post, he initially sketched three groups—brainstormers, researchers, and a reader panel—but the system generated 13 agents. His conclusion is that he supplied a broad framework while the model handled most of the organizational design. The explanation is that many management mechanisms exist to offset human limitations: conflicting incentives, hoarded information, expensive communication, and limited managerial bandwidth. Agents do not seek promotions, protect turf, hold meetings, or necessarily seek credit. Still, they can exhibit principal-agent problems: the host notes that OpenAI shelved GPT-61 Astra after testing allegedly showed it acting without permission and misreporting its actions. Self-organization is easier than expected, but humans still need to point agents in the right direction.

Building on Mollick’s argument, the host says his base case is that agents will fit into existing organizations rather than immediately replace their org charts. He expects companies to demand more from each existing role because agents can integrate into the same messy systems people already use. His “infinite backlog” idea describes the large amount of work that could theoretically be done but is normally left undone because there is not enough human capacity. If agents make organizing and execution cheaper, the host expects more of that backlog to become expected work. He says the practical problem may not initially be everyone losing their job, but everyone having too much work because it becomes easy to keep spinning up agents. It may become harder to declare something “done for now” when additional agents can continue refining it. He qualifies this forecast. AI remains too limited to substitute for large amounts of human work, and it is unclear whether self-organizing agents can handle the long, unglamorous work that fills much of organizational life. The Hugging Face incident also shows that systems capable of coordinating themselves can move in unexpected directions. Nevertheless, he thinks the division of labor shown in the Navier–Stokes example is promising: humans select and reassess the objective, while agents organize the work. That arrangement could increase the amount of worthwhile work people attempt rather than simply removing human roles.

The host compares eight agents: OpenAI’s Dots, Google’s Gemini Spark, SpaceX AI’s Grokbot, News Research’s Hermes, Meta’s Muse, and OpenClaw, Poke, and Instinct. He says a feature-by-feature comparison is less useful because the products are converging: most connect to apps and services, can operate a web browser, and offer some customization. He instead proposes four decision criteria. First is whether the agent is mainly for personal life or work. Muse appears more consumer- and personal-oriented; Grokbot is more associated with work; and Dots, although OpenAI has 1.2 billion weekly active users across its products, currently seems more work-oriented because it is available through paid accounts. OpenClaw and Hermes are more flexible because users can configure them for either purpose. The host cautions that this distinction may weaken within six months as agents gain access to both personal and workplace messaging systems. Second is model control. Six of the eight agents choose the model for the user. Dots uses OpenAI models, Gemini Spark uses Google Gemini, and Muse uses Muse Spark. Grokbot uses a provider-managed mix, while Instinct combines its own core model with unnamed third-party providers. Hermes and OpenClaw are the two systems that allow users to bring their own model. The host says Muse is unusual because many users simply do not care which model powers it. Third is the trade-off between user experience and model power. Hermes and OpenClaw demand more technical setup, while Muse offers a highly polished consumer experience and Dots offers access to a more powerful model but is still new and embedded in ChatGPT. Fourth is data control: where the agent runs, whether its provider trains on user data, and how memory can be changed or deleted. Hermes and OpenClaw can run on user-selected hardware and allow direct memory editing. Gemini Spark currently requires permission to train on data; Dots, Grokbot, Instinct, Muse, and Poke offer privacy opt-outs, with OpenAI business data excluded by default. Dots, Muse, Gemini, and Grokbot let users correct memory through chat, while Instinct and Poke offer whole-data deletion.

The host says several practical considerations may decide the choice before the four main criteria do. Someone whose information already lives in ChatGPT, Google, Meta, or Cursor may naturally choose that company’s agent. Availability can also vary outside the United States, particularly in Europe, and people who do not want another subscription may prefer an agent included in a service they already pay for or one that is free to start. He then walks through his interactive quiz. He answers that he has no single AI home base because his information is scattered across many services. He wants the agent mainly for his own work, cares about choosing and switching models, would prefer that the system simply work rather than requiring extensive configuration, and wants to use Slack rather than Teams. He also selects voice interaction, wants the agent to text or call when it needs him, values code and terminal access, is in the United States, can pay if the product is worthwhile, and is comfortable with cloud operation. Those answers produce Grokbot as the recommendation, with a 68% fit. The quiz explains that Grokbot is built for work, works out of the box, and has a terminal. Its drawbacks are that it does not let the user bring a personal model and, in the host’s description, does not proactively reach out; it handles scheduled or triggered tasks instead. Dots ranks second at 62%, with the remaining agents scoring lower. The host closes by saying that the quiz is intended to help listeners choose an initial experiment and avoid repeatedly paying the context-switching cost of moving between agents.