Persistent AI Agents: How to Get Started (and What to Watch Out For)

Host Michael Stelzner asks Jonathan Mast for the biggest misconception about AI agents. Mast says people imagine handing over the business, going to the beach with a Mai Tai, and letting AI do everything; although that might eventually be possible, it is nowhere near reality. AI amplifies a person’s skills and experience, so the human remains necessary. Mast says the practical upside is different: agents perform heavy-lifting and repeatable work, freeing time for important personal and business tasks that people continually postpone.

Mast illustrates the benefit with his own subscriptions. His agent examines his email and each morning reports renewals that arrived overnight, asking whether he still wants them. If Mast says no, the agent researches the cancellation process, locates the often-hidden cancellation button several layers into a website, and submits the cancellation with his approval. Mast says he had avoided canceling memberships because the process could take half an hour, whereas the agent now performs that work. The example demonstrates that persistent agents can act beyond a chat response when authorized.

Before deploying an agent, Mast recommends treating it as a virtual employee and completing “agent onboarding.” Define privacy and security rules first: whether it may send email, delete files, or access sensitive information without approval. A new agent should not receive the equivalent of keys to the company safe on its first day. Also specify where its work belongs; Mast’s company uses Google Drive, so agents receive access to particular folders and instructions about placement and formatting. Finally, establish when it may control a local computer, since an agent taking over the keyboard during active work can be disruptive.

Mast says the defining difference from earlier chat tools and ordinary agents is persistent memory. Instead of starting a new conversation that forgets prior instructions and requires files to be uploaded again, a persistent agent can remember how Mast works and retrieve what he told it—he even asks it what they discussed the previous week. The interaction is one continuing chat chain rather than separate conversations. Communication can happen through text messages, WhatsApp, Slack, Discord, or Telegram; Mast’s Instinct agent texts him like a human contact.

Persistent agents can continue an assignment in the background through cloud services after the user closes the computer. A conversation started on a desktop can continue from a phone or tablet during a meeting, dinner, or vacation. Mast contrasts this with earlier systems such as OpenClaw and Hermes, which often required a dedicated computer or virtual private server and ongoing technical maintenance. He names GrokBot, Dots, Muse, and Instinct as newer services that largely “just work,” requesting permissions through ordinary authorization screens instead of command-line setup.

Mast has given persistent agents to himself, his chief of staff, CEO, customer-service leader, and other employees so they can hand work to one another. For example, his agent can tell the CEO’s agent that an item is ready for review, following a workflow Mast previously defined. During a difficult membership-platform migration involving PayPal, Stripe, and a new management system, the agents compare renewals, new subscriptions, and payment data and assemble a shared dashboard. Mast says this eliminated daily report-pulling, meetings, and developer costs for building and maintaining the dashboard.

Mast says his agent helps because he was diagnosed with ADHD about 13 years ago and often forgets commitments once they are out of sight. Connected to voice recordings, meeting transcripts, calendar, and task lists, it can warn him during a meeting that he promised something for a day when he will be out of town. It also noticed that Mast kept adding Google Tasks without completing any, so it now spends 10–15 minutes with him each day reviewing open tasks and keeps reminding him when he ignores the review. Mast compares it to an experienced executive assistant.

For security, Mast recommends an agent that assumes it needs permission rather than one that can act unless stopped. He forbids his agents from buying anything or deleting files without approval and limits their Google Drive access to a shared folder. His preferred rule is “ask before acting”: the agent may research, plan, and prepare without permission, but it must ask before sending messages, making commitments, or contacting outsiders. After Mast mentioned roof damage from a hurricane, the agent researched five local roofers and offered to request quotes, but stopped until he approved outreach.

Mast exported comprehensive profiles from ChatGPT and Claude—one document was 30–40 pages and another about 50—and gave them to his new agent. The material covered his ADHD, family, employees, roles, email addresses, phone numbers, company, products, and services. This allowed the agent to route a membership customer’s failed coupon renewal to the correct customer-service leader, Sarah, after Mast approved it. Mast also supplies DISC and Myers-Briggs information. The agent can then adapt one message for a direct, high-level founder and another for a detail-oriented HR director, reducing generic “AI slop.”

Mast names four comparable services: GrokBot from xAI, Muse from Meta, Dots from OpenAI, and Instinct, whose backers he says are less clear but which is a Wall Street favorite. Muse and Instinct were free when recorded; Dots required a business or enterprise plan costing Mast’s company $125 per person monthly, with five to ten times more usage, and his leadership team chose it because employees already knew ChatGPT. He recommends beginning with calendar and email access, while prohibiting unsupervised scheduling or sending. The agent can triage messages, organize old email, flag missed items, and route work to teammates.