Skip to content
Workprentice

Introduction

Workprentice gives your team AI agents that take recurring work off your plate. An agent connects to the tools your team already uses, does a job when you ask or on a schedule, and hands back a finished deliverable, such as a document or a spreadsheet, together with a record of every step it took. Think of each one as an apprentice: it learns how your team works, remembers what you correct, and gets more useful the longer it works for you.

A concrete picture helps. Suppose your team wants a sales-pipeline report every Monday morning. Today someone spends an hour pulling CRM data, comparing the numbers to last week’s, and formatting a summary. With Workprentice you connect the CRM once, describe the report once — “a document with a summary table, the top five at-risk deals, and week-over-week changes” — and set a weekly schedule. From then on the report arrives as a finished document, and each run leaves a complete record of how it was made: what the agent queried, how it reasoned, and what it produced. Most of the product follows this pattern: you describe a job once, then spend your time reviewing results.

That is the main difference from a chat assistant, which answers while you sit with it and leaves you to carry the result where it’s needed. An agent keeps working after you leave, reaches into the systems where the data lives, and remembers the correction you made last week.

The shape of the product

Section titled “The shape of the product”
  • Agents do the work. Your personal agent is ready the moment you sign up, pinned at the top of the sidebar: teach it your preferences, ask it questions, hand it one-off jobs. You can also create agents for specific jobs, such as a support triager or a pipeline analyst, each with its own personality, skills, connected tools, and limits. A new agent starts private to you, so you can shape it before sharing it with the workspace.
  • Assignments are jobs you describe once. Run them by hand, put them on a schedule (hourly through monthly, or a custom cadence), or let another system start them with a webhook. If you find yourself asking an agent for the same thing repeatedly, Save as Assignment turns that conversation into a draft assignment.
  • Sessions are the record of each conversation or run. Every step the agent takes is visible, live while it works and afterward for good.
  • Work products are what a session produces: spreadsheets, PDFs, drafts, charts, images, videos, and links. Each one links back to the session that made it, and all of them collect on the Work Products page. Work meant to last is kept as a document, data table, or app.
  • The Inbox is where agents reach you. When an agent hits a blocker, such as expired credentials, missing data, or an ambiguous instruction, it sends you a message with a priority instead of guessing or failing silently. The message can follow you to Slack, email, or text; you reply, and the work continues from there.

Around those five sit the capabilities that give agents their reach. Integrations connect them to your systems, such as CRMs, ticketing, chat, and code, through the open Model Context Protocol. Skills carry reusable know-how, such as a report format or the way your team qualifies deals. Memory lets an agent keep what it learns across conversations, and knowledge bases hold the shared reference material (handbooks, playbooks, style guides) that agents consult while they work.

How you stay in charge

Section titled “How you stay in charge”

An agent that works while you’re away is only useful if you can check what it did and set limits on what it may do. Both are built in.

Sessions are the audit trail. Every step the agent takes is written into the session and stays there, so “what did the agent actually do?” is answered by opening the session and reading it. While a session runs, you can watch it, stop it, or send a message to redirect it without waiting for it to finish. Checking your agent’s work shows how.

An agent can only use the tools you have connected, and guardrails narrow that further: which integrations, shell commands, and network destinations each agent may use, whether it may work with other agents or refine its own skills and personality, and how much a single session may spend before it pauses for your review. The same controls exist at the organization, workspace, and agent levels, and a lower level can tighten what a higher one set but never loosen it, so you can set a floor once for the whole company and tune individual agents from there. Guardian, on by default, reviews what an agent is about to push or post for confidential information before it leaves.

Next steps

Section titled “Next steps”

For the complete site index, see llms.txt.