Berman begins with email triage. A scheduled agent groups messages that can be archived, summarizes low-effort items and prepares replies for messages that need action. Connected context from Gmail, Drive and business systemsContext engineering designs the information, instructions, memory, and tool state an AI receives so it can perform a task reliably. helps it classify sponsorships, retrieve prior conversations and suggest the next step.
Separate bots manage calendars, browser tasks and coding work. Calendar automation negotiates meeting times and creates events from email, while browser control handles repetitive website flows. A coding bot can delegate to Cursor, monitor pull requests and continue until automated checks passAgent completion verification checks observable evidence that an agent achieved the requested result instead of accepting its claim that the work is finished..
Meeting transcripts become summaries and action items on a recurring schedule. Berman proposes letting the agent compare those commitments with an existing task list and complete suitable items automatically. A chief-of-staff bot provides one front doorA coordinator agent interprets goals, assigns work to other agents, tracks dependencies, and combines or escalates their results. and delegates to narrower specialistsA specialized AI agent is configured for a focused role with task-specific context, tools, models, and permissions. for work, food, school and personal administration.
He also demonstrates conservative computer cleanup, where the agent only proposes files by risk level until the user approves deletionHuman-in-the-loop command approval requires a person to review and authorize a consequential AI action before it executes., plus Telegram access for voice notes and remote coordination. The recurring theme is narrow responsibilities, persistent context and explicit confirmation before consequential actions.
Watch on YouTube
