ワークスペースエージェント

Most ChatGPT users already know how to use AI for one-off tasks—like drafting, summarizing, brainstorming, or answering questions. The next phase of AI use is broader and more embedded in day-to-day work. Instead of helping with isolated moments, AI is increasingly being used to support repeatable workflows that depend on shared systems, standard handoffs, consistent outputs, and real-world constraints like timing, accuracy, and process.That’s where workspace agents in ChatGPT fit. They’re designed to be used for repeatable workflows—work you’d otherwise do manually, re-explaining the steps each time, and copying information between tools. Learn more about workspace agents in our blog post.If you’re new to agent building, let’s focus on the core concepts first so when you start building, you’ll know how to set up your workspace agent for consistent results.Generally speaking, an agent is a system that carries out a task with three components: a trigger, a process that may include specialized skills, and tools or systems it can connect to.ComponentDescriptionExamplesTriggerWhat starts the agentA schedule (“Every weekday at 9am”) or a manual run (“Run now”).Process and skillsThe steps the agent follows to complete the task the way you expectReviewing inputs, checking for missing information, drafting an output, and handing it off or taking the next action.Tools and systemsThe approved tools and integrations the agent can use to gather information and, if allowed, take actions.Slack, a CRM, internal documentation, a ticketing system, or a shared documentAgents are most useful when the work is:Repeatable: The same task comes up regularlyStructured: There’s a clear format for the output (so you can tell if the agent is doing a good job)Time-based or event-driven: It runs on a cadence or is triggered by an eventTool-based: It requires reading from or writing to systems your team usesFor open-ended thinking, brainstorming, or exploratory writing, regular chat is often a better fit—especially for one-off tasks.Agents are also different from traditional API workflows you may have built in the past. Traditional workflows in other tools are often deterministic, meaning each step is explicitly defined, and the system follows the same path each time unless you change the logic. Agents are more probabilistic. They still operate within instructions, tools, and guardrails, but they use a model to interpret context, make bounded decisions, and adjust how they move through the work.A helpful way to design a workspace agent is to break it into parts. Think about what you would clarify before handing work to a person: what they are responsible for, when they should begin, what should make them pause or stop, which tools and information they can use, the process they should follow, and the rules they must stay within. An agent may require access to apps. Learn more about apps in ChatGPT.⁠(opens in a new window)See some examples of agent breakdowns below:Marketing campaign summary agentProduct feedback triage agentSales pipeline summary agentObjective: What is the overall goal of the agent?Analyze campaign performance and recommend optimizationsSummarize new product feedback and route to the right ownersMonitor pipeline changes and highlight risks and opportunitiesTrigger: What starts the work?Every Monday at 10 AMForm submission from SlackEvery weekday at 8 AMProcess: What steps should it follow? Pull KPIs, identify trends, draft a summary, propose next steps, assign ownersReview feedback, group related items, summarize key themes, suggest owners for follow-upReview pipeline, flag risks, propose next stepsTools: What systems or applications do you need access to?Analytics tools and a shared summary documentSlack, ticketing system, and a summary documentCRM, email or Slack, and a pipeline trackerGovernance: What controls do you need? When should it stop and escalate?Draft recommendations; approval required before budget changesCreate draft tickets; escalate high-priority issues; approval required before submissionDraft insights and recommendations; notify owners; no direct outreach without approvalThe examples below can be thought of as agent workflow patterns: common, repeatable patterns of work that show up across teams and functions. Each one represents a core type of workflow an agent can carry out. The specific tools, data, and outputs may vary, but the underlying pattern stays consistent.WorkflowHow it worksExample use casesBriefingPull information from multiple places, distill it, and package it into something decision-ready.1) Collect inputs2) Compare and extract key signals3) Summarize for audience 4) Share as doc, memo, or briefingSales: Build account briefing from CRM, calls, Slack, and news. Marketing: Compile campaign or competitor recaps from analytics, social, and docs. Exec: Produce daily industry briefs with recommendationsTriage and routingProcess inbound items and ensure they reach the right next step1) Review inbound items 2) Categorize and prioritize 3) Route or create next-step artifact4) Notify owner and requesterSupport: Turn feedback into bugs, feature requests, or follow-ups. IT/Ops: Handle internal requests in Slack and route by urgency. Recruiting: Screen inbound candidates and move them into the right path.Analysis and recommendationInterpret data or evidence, form a point of view, and turn that into a first deliverable.1) Pull source data 2) Analyze patterns, gaps, and tradeoffs 3) Form recommendation 4) Draft memo, deck, spreadsheet, or emailFinance: Reconcile budget vs. actuals and draft a manager memo. Product/Research: Analyze user feedback and propose priorities. Procurement: Compare vendor quotes, create decision matrix, and draft PO email.Content creationCreate or update content, then tailor it for a specific audience or channel.1) Generate first draft from notes or source material 2) Edit for tone and accuracy 3) Tailor to audience and channel 4) Publish or sendMarketing: Turn a brief into campaign assets and circulate for feedback. Manager: Turn notes into onboarding plan or training module. Sales: Write follow-up emails, QBR summaries, or stakeholder recaps.Planning and coordinationTurn goals into scheduled work and system updates1) Build plan from constraints 2) Check dependencies and availability 3) Take action or prepare action 4) Update systems and notify peoplePM/Chief of Staff: Block work, schedule meetings, and update trackers. Events: Coordinate off-sites, registrations, or travel plans. Admin: Manage forms, reminders, orders, bookings, and follow-ups.You may start by using agents your organization has already built. Begin by understanding what the agent is designed to do well—what tasks it supports, which tools it relies on, and the kind of output it produces.Start with a few low-risk requests to see how it behaves. Try simple inputs first, then review the results to understand how it approaches the task.Keep in mind that even a well-built agent still benefits from human judgment. You are often the one who knows the broader context, the stakes of the task, and what a good answer should look like.Once you understand the basics and have identified a strong use case, you can start building your own workspace agent. You may even already have skills built in ChatGPT. Workspace agents can use skills in their instructions.Note: in ChatGPT Enterprise, access to build agents is controlled by your workspace administrators.Start in plain language: In the agent builder chat, describe the job the agent should do, what a successful outcome looks like, and any constraints it needs to follow. The builder can translate this into a clear workflow with defined steps, which you can keep refining in chat or edit directly in the workflow and instructions.Choose tools and connectors: Select the approved apps the agent can use to complete the workflow. You can also describe the systems the agent needs to access, and the builder will guide you through adding and authenticating them.Choose a trigger: Decide when the agent should run. An agent trigger is the event that starts the agent running. For workspace agents, it can be human-triggered (someone asks it to do something), schedule-triggered (it runs at a set time). You can use the agent builder chat to set up your trigger using plain language.Add guardrails: Set boundaries, required approvals, and any human-in-the-loop checkpoints for sensitive actions. You can add these directly into the agent builder chat, and the agent builder will modify the instructions.Agent building works best as an iterative process. In ChatGPT, the builder is part of that loop. As you test your agent, use the conversation to understand what happened, spot issues, and refine the instructions step-by-step.Start with a few realistic examples, including both straightforward requests and messier ones with missing context or ambiguity. This helps you see how the agent behaves and where it may need clearer instructions or guardrails.As you test your agent, it’s expected that the first version won’t be perfect. When something feels off, there are two effective ways to improve it.Update the instructions directly: Make small, specific changes in the editor, such as clarifying a step, adjusting the output format, or adding a constraint.Coach it in natural language: When the issue is less clear, use the builder conversation to work through it. You can point out what it missed, ask open-ended questions, or have it explain its reasoning to surface gaps in the instructions.After making changes, test again to confirm the update worked.Workspace agents are designed for shared, repeatable work. Sharing an agent gives your team a consistent way to complete a task instead of reinventing the process each time.When you share an agent, be explicit about what it’s for. In the description, include the task it handles, when to use it, what inputs to provide, and what kind of output to expect. Shared agents work best when they’re tied to a specific, recurring workflow your team already understands. Including one or two example prompts can also give others a clear starting point and make adoption smoother.Remember that workspace admins manage connector and feature access through role-based access control (RBAC), so teammates may need the appropriate permissions for the agent to work with systems like Slack, Gmail, or other tools.