Talavine
Pick a role
Start from a blueprint like Technical Support or Marketing Outreach.
Agent employees
Put a helper on a job that keeps coming back.
Set up an agent to collect information, research a subject, or prepare a customer follow-up. Your data hub supplies the relevant records; you choose the sources, tools, schedule, and permissions for the work.
Named roles
Easy setup
Boss approval
Visible activity
Describe the job
Describe the result you want, the information to use, and when the helper should ask you for help.
Let setup prepare the tools
Setup can prepare the mailbox and connected services described in the job. You supply the required accounts and access.
Set the send policy
Choose drafts for your review, or allow direct replies for specific kinds of work.
Real jobs, not mascot prompts
The point is not more chat. The point is work getting handled.
Roles carry the job brief, the information sources, the setup instructions, and the boundaries. These examples are the ones we're using to make the system practical.
Executive Assistant Employee
Works the inbox on a schedule so urgent things surface, follow-ups get drafted, and nothing slips while you're in meetings.
Triage a mailbox on its schedule: separate the urgent items, pending replies, and calendar changes from the noise.
Draft internal replies and follow-ups for threads that have gone quiet, ready for your review.
Spot meeting conflicts and prepare alternate-time options instead of letting them collide.
Send you a concise report each cycle that separates facts, proposed actions, and blockers.
Tech note, translated: The helper checks its assigned mailbox on a schedule and prepares drafts and proposed next steps. Its activity log shows what it reviewed and produced.
Boss rule: It starts draft-first. You approve what goes out, and any feedback you leave on a proposal can become a standing rule the employee follows from then on.
Technical Support Employee
Answers software questions by checking how the product actually works, then replying from the support inbox when policy allows.
Watch a support mailbox such as support@example.com.
Read the customer's question and identify the real job they are trying to do.
Use a local repo, GitHub MCP, docs, and app screenshots as source-of-truth context.
Run or inspect the app when it needs to prove the answer instead of guessing.
Send the answer from the support mailbox when the answer is grounded, safe, and inside the role policy.
Tech note, translated: The role is a setup blueprint, not just a prompt. It can carry mailbox details, local repo paths, GitHub context, screenshots, escalation rules, and a repair loop for missing tools. The employee should know how to get ready before the first customer message arrives.
Boss rule: For test or low-risk support, you can allow direct replies. The host still keeps receipts and the employee escalates when the answer is not grounded.
Marketing Outreach Employee
Finds podcasters, YouTubers, and writers who already care about agentic AI, then works the conversation until a real booking needs you.
Research public episodes, channels, posts, and contact pages for people already discussing agentic AI systems.
Build a source-backed outreach list with why each person is a fit.
Draft specific outreach that references the person's actual work instead of a generic pitch.
Manage replies until there is a date, recording notes, follow-ups, and handoff points for the boss.
Tech note, translated: Public web research, mailbox access, calendar awareness, artifacts, and source receipts combine into one job. The role can require verified contact routes and forbid invented addresses or fake engagement claims.
Boss rule: You decide how far it can go. Many teams will keep first contact approval-gated, then allow routine scheduling replies once a person has engaged.
Boss / employee model
You give the employee a job, not a blank check.
That's the important difference. The employee has a role, allowed tools, information sources, repair instructions, and reporting rules. It can work in the background, but boss policy still controls public commitments, deleting, spending, and broad policy changes.
Boss
sets role, tools, approvals
Employee
sets up, watches, researches, replies
Report
draft, proposal, or result
What the boss controls
The safety rails are part of the employee's job description.
The job description: what this employee is for and what it should ignore.
The approved systems it may use, such as mail, search, calendar, or GitHub MCP.
Model and spend limits so a background job can't quietly run up cost.
The line between 'prepare it' and 'go do it.'
Progress notes and proposals that show up for the boss to review.
If credentials, tools, or network access break, the employee explains the fix.
Employees stay visible.
You can see the roster, status, role, and where each employee is blocked or waiting.
The work becomes trackable.
Research, follow-ups, support replies, and approvals can turn into tasks instead of disappearing back into the inbox.
Start with one employee and a narrow job.
Give it one inbox, one role, and clear policy. Once you trust the loop, add more sources and more responsibility. Watch one get built →