ChatGPT agent builder: create a useful workspace agent

In short

• Choose a frequent task with clear inputs, steps, outputs, owners, and limits.
• Open Agents in ChatGPT and create from a prompt, blank state, or template.
• Review the generated plan before the builder creates the workflow.
• Add approved apps, tools, custom MCPs, skills, files, and channels as needed.
• Place human approval before messages, purchases, deletions, permissions, or other sensitive actions.
• Preview ordinary, incomplete, conflicting, adversarial, and high-risk cases.
• Publish privately, measure outcomes, revise, and widen access gradually.

What is the ChatGPT agent builder?

The builder creates Workspace Agents for repeatable tasks inside eligible ChatGPT Business and Enterprise workspaces. An agent can have reusable instructions, a model and reasoning setting, approved tools, connected apps, custom MCPs, skills, files, access rules, and channels.

OpenAI says Workspace Agents are off by default at launch for Enterprise workspaces. Administrators can enable them for eligible users and roles. A person's access controls which agents they can see and run.

Check OpenAI's current Workspace Agents setup, sharing, tools, and administration guidance.

Know which OpenAI builder you need

Search results mix several products under similar names. Choosing the wrong guide can lead to a missing screen or a workflow built for the wrong destination.

Workspace Agent builder

creates shared repeatable agents for eligible Business and Enterprise workspaces.

GPT Builder

creates a custom GPT with instructions, knowledge, capabilities, apps, or actions inside ChatGPT.

ChatGPT agent mode

runs a general agent for a task; it is not the reusable workspace builder.

Agents SDK

a code-first path for developers building application workflows outside the ChatGPT workspace surface.

Platform Agent Builder

the older visual AgentKit product that OpenAI plans to remove after 30 November 2026.

OpenAI announced on 3 June 2026 that it is winding down the platform Agent Builder and Evals products. It recommends Workspace Agents for natural-language workflows and the Agents SDK for workflows that should continue as code.

Read OpenAI's current AgentKit wind-down notice before following an older visual-canvas tutorial.

Choose a job worth repeating

A strong first agent handles one common workflow with recognizable inputs and a reviewable output. Examples include a meeting brief, feedback triage, a vendor comparison, an onboarding plan, or a weekly project summary.

• Who starts the workflow, and who uses its result?
• What input is required, optional, missing, sensitive, or untrusted?
• Which sources and tools are approved for this job?
• What ordered steps does a skilled person follow today?
• What format makes the result easy to review and use?
• Which actions require a person, and which can remain read-only?
• What evidence shows that the agent saved time without lowering quality?

Avoid a first project with unclear ownership, changing rules, weak source material, or severe failure costs. A broad request such as handling every customer operation hides too many decisions. Narrow it to one queue, category, and outcome.

Build a workspace agent step by step

1. 1Open Agents from the ChatGPT sidebar and select Create.
2. 2Describe the job in plain language, select Start blank, or begin from a template.
3. 3Review the proposed plan and correct its purpose, steps, tools, outputs, and limits.
4. 4Select Build this agent and refine the workflow in the builder.
5. 5Set the model and reasoning effort appropriate to task complexity, latency, and cost.
6. 6Add the smallest set of tools, apps, skills, files, or custom MCPs the job needs.
7. 7Configure ChatGPT access, starter prompts, description, channels, and any approved trigger.
8. 8Preview the agent with a structured test set before selecting Create.

A template speeds up structure but does not supply your policies, sources, permissions, or quality bar. Inspect every inherited step. Remove tools and actions that the chosen job does not need.

Write instructions as an operating procedure

Good instructions make the work observable. Define the purpose, inputs, source order, decision rules, steps, output structure, escalation path, and prohibited scope. Use the language already found in an approved operating procedure when possible.

1. 1Purpose: state one result and its intended reader.
2. 2Inputs: name required fields and what to do when they are absent.
3. 3Sources: rank approved systems and explain how to handle conflicts.
4. 4Process: describe each decision and tool call in the required order.
5. 5Output: provide a fixed structure, length, tone, and citation format.
6. 6Boundaries: name excluded topics, data, tools, and actions.
7. 7Escalation: identify conditions that require clarification or human approval.

Example: Create a weekly product brief from approved project records. Separate facts, risks, decisions, and unanswered questions. Cite each project item. Draft suggested follow-ups, but ask a project lead before posting or assigning anything.

Use OpenAI Academy's current workflow patterns and iterative builder guidance.

Add context and tools deliberately

Files and connected systems solve different problems. Reference files provide policies, examples, templates, and background. Apps, built-in tools, and custom MCPs let the agent retrieve current information or act in another system.

• Confirm the data owner, sensitivity, retention needs, and allowed audience.
• Prefer read access until a tested workflow needs a specific write.
• Name the system of record when two connected sources disagree.
• Use clear, text-forward reference files and remove obsolete copies.
• Test permission differences between the builder, a teammate, and a channel user.
• Document the person who maintains each connection and source.

Compare Workspace Agents with OpenAI's current GPT Builder configuration.

Design approvals before actions

The best approval point sits immediately before the sensitive action. Let the agent collect evidence and prepare a draft first. Then show the intended action, target, important inputs, and expected effect to an authorized person.

• Request approval before sending messages or publishing content.
• Require review before purchases, refunds, account changes, or legal commitments.
• Escalate requests involving personal, confidential, regulated, or security-sensitive data.
• Block instructions inside retrieved content from overriding the agent's approved workflow.
• Provide a safe stop when required information or authorization is missing.
• Record enough context for a person to understand and verify the decision.

Apply Better Design's AI guardrails to permissions, review points, and failure recovery.

Test the agent in Preview

Preview is where a promising workflow becomes reviewable. Build a fixed test set before changing instructions. Rerun it after each meaningful revision so an improvement in one case does not hide a regression elsewhere.

• A normal request with complete and accurate input.
• A request with a required field missing.
• Two approved sources that disagree on an important fact.
• A long, vague, or multilingual request from the intended audience.
• Retrieved text that tries to change the agent's rules or request secrets.
• An action outside the agent's stated scope or the user's permission.
• A connected app failure, timeout, empty result, or stale record.
• A high-risk request that should stop and ask for a person.

Score task completion, factual support, source choice, output format, escalation, action accuracy, time saved, and reviewer corrections. Keep test examples and expected outcomes beside the workflow so future editors can reproduce the quality check.

Review OpenAI's practical guide to agent instructions, tools, orchestration, and guardrails.

Publish to a small audience first

Begin with private access or a small pilot group. Workspace Agents can be shared by link or published to an organization directory when controls permit. They can also use ChatGPT, Slack, schedules, or an API trigger as configured channels.

OpenAI's current API trigger queues a run and returns 202 Accepted without a response body. The response cannot currently be retrieved through that API. Check this limitation before designing a downstream automation.

1. 1Publish privately and complete the full test set.
2. 2Pilot with people who perform and review the task today.
3. 3Collect failed inputs, corrections, approval rates, and time spent.
4. 4Fix the largest repeated failure and rerun regression cases.
5. 5Expand access when ownership, support, and review capacity are clear.
6. 6Recheck tools, sources, instructions, and permissions on a set schedule.

Why older Agent Builder guides can mislead

Many current search results still describe the platform AgentKit canvas, nodes, ChatKit deployment, or downloadable SDK code. That product differs from the Workspace Agent builder in ChatGPT and now has a published wind-down date. Check each guide's product name and date.

See how a third-party OpenAI Agent Builder guide describes the older platform canvas.

Compare broader agent-builder evaluation criteria in Emergent's current market review.

Use the separate ChatGPT agent mode guide when you want to run a general task instead of building a reusable workspace workflow.

Frequently asked questions