Custom GPT Development, Migration and Workspace Agents
A custom GPT is a version of ChatGPT configured for one job: instructions, knowledge files and, where allowed, apps or Actions, shared inside your workspace. As of October 2026, OpenAI has scheduled custom GPTs to retire on December 11, 2026 and recommends plugins; the newer builder inside ChatGPT Business (formerly ChatGPT Team) and ChatGPT Enterprise is workspace agents. NetSys builds what companies used to build as custom GPTs on those successors, migrates the GPTs you already rely on, and tells you when an API-built agent is the better fit.
The short answer
Custom GPT development means configuring a version of ChatGPT for one business job: the instructions, the approved knowledge, the systems it may call, who may use it and how it is tested and maintained. OpenAI has scheduled custom GPTs to retire on December 11, 2026 (as of October 2026), with plugins and workspace agents as the successors, so NetSys scopes each build on the current tools, migrates existing GPTs (knowledge files carry over as plugin reference files; GPT Actions do not), and recommends an API-built agent when the job needs to run outside ChatGPT.
Closest related page: AI agents for business: consulting and development. That page covers agents built in Copilot Studio or on Claude that act in your systems with approvals; this page covers assistants configured inside your ChatGPT workspace, and when to move from one to the other.
GPTs (custom GPTs) combine instructions, conversation starters, knowledge files, capabilities such as web search and code interpreter, and either connected apps or Actions, not both. They can be built and published in Business, Enterprise and Edu workspaces when the settings allow, shared by invitation, by workspace link, with the whole workspace or through the GPT Store, and builders cannot see the conversations people have with them. Personal plans can no longer create or publish them.
OpenAI's retirement schedule, as published in October 2026: creation of new GPTs on Enterprise is planned to end on October 26, 2026; the standard retirement date is December 11, 2026; Enterprise workspaces with an approved deferral have until February 11, 2027. The recommended successor is a plugin (skills plus apps or a custom MCP server). Knowledge files migrate as plugin reference files, GPT Actions do not transfer, and the migrated plugin starts private. Workspace agents are the newer builder for Business and Enterprise: a chosen model and reasoning effort, tools, apps, custom MCP servers, skills, files, schedules, a Slack channel and an API trigger, with write actions set to always ask by default.
So custom GPT development in late 2026 is three jobs: migrating or rebuilding the GPTs a company depends on, building new assistants as workspace agents or plugins, and knowing when the job belongs outside ChatGPT in an API-built agent with its own identity, approvals and logs. This page covers all three.
One job, approved knowledge, a test set, an owner
Discovery picks one job that staff do often enough to matter and writes down the inputs, the approved sources, the actions (if any) and the person who owns the result. We draft the instructions and assemble the knowledge from documents you have approved, then test against a representative set of real questions, with pass criteria agreed before anyone sees a demo. Sharing is set to the smallest audience that needs it, the admin controls are checked, and a one-page employee guide explains what the assistant is for and what it is not. If the job needs to write into a system, we build it as a workspace agent or a plugin backed by a custom MCP server with approvals, or move it to an API-built agent. After launch it is tested every month under managed AI services.
What custom GPT and workspace agent development includes
Scope and fit
A GPT, a shared Project, a workspace agent or an API agent, chosen by the job.
- One recurring job with an owner, written as a one-page brief: inputs, sources, actions, audience, pass criteria
- The fit decision: static curated knowledge (a plugin, which replaces the GPT), a living team context (a shared Project), scheduled or triggered work (a workspace agent) or work outside ChatGPT (an API-built agent)
- A plan check: what your Business or Enterprise workspace allows for building, publishing and sharing
- A migration plan for any GPT you already rely on, ahead of the December 11, 2026 retirement
Knowledge and instructions
Approved documents in, tested instructions on top.
- Knowledge assembled from documents a named owner has approved, with outdated versions removed first
- Instructions written for the job: tone, which sources to prefer, what to refuse, when to hand off to a person
- File limits and formats checked against OpenAI's current rules for GPT knowledge, plugin reference files and agent files
- Instructions and knowledge versioned, so a change can be rolled back
Apps, Actions and custom MCP servers
Read first; write only with an approval step.
- Connected apps (SharePoint, Google Drive, Slack and others) used for read access inside each user's existing permissions
- GPT Actions (an OpenAPI schema plus an API key or OAuth) only where a short-lived GPT still makes sense, because Actions do not transfer at migration
- Custom MCP servers for write actions, which on Business only owners and admins can create and publish, with the tool list refreshed after each change
- Workspace agent write actions left on always ask; agent-owned versus end-user authentication decided per job
Sharing, permissions and admin controls
The smallest audience that needs it.
- Sharing set per item: invite-only, workspace link, the whole workspace or, rarely, the GPT Store
- Permission levels (can chat, can view settings, can edit) assigned to named people
- On Enterprise, role-based access for who may create, publish and share, and the action domain allowlist maintained
- Ownership recorded, with transfer handled when the builder leaves
Representative evaluation and employee guide
Tested on real questions before anyone depends on it.
- A test set of real questions and documents, with expected answers agreed by the owner
- Pass criteria for accuracy, refusals and handoffs, re-run after every change
- A one-page employee guide: what it is for, what it must not be used for, how to report a wrong answer
- Prompt injection checks on uploaded documents and connected sources before launch
Maintenance and migration
Assistants drift when the business changes.
- A monthly test-set run with instruction and knowledge updates under managed AI services
- GPT-to-plugin migration run with the owner: knowledge carried over, Actions rebuilt as MCP tools where still needed, sharing reset deliberately
- Retirement of assistants whose owner has left or whose job no longer exists
- A register of every published GPT, agent and plugin with its owner, audience and last test date
Why businesses have NetSys build their GPTs and agents
Tell us the job you want an assistant to do, which documents it should know, whether it needs to write into any system and who would use it. A NetSys engineer will say whether it belongs in a plugin, a workspace agent or an API-built agent, and scope a pilot.
Who does the work: meet the NetSys team on our About page.
- Built on the current tools and the published retirement dates, so you do not pay for an assistant that stops working in December
- The same method as our published AI agent work: a written charter, least-privilege access, approvals on anything that writes, and a test set before launch
- Delivered work: custom agents on Copilot for an 18-person organization, Claude apps for a lean team, and a Copilot and Claude rollout for 25 employees
- The workspace is yours: every GPT, agent and plugin lives in your account under your owners
- Plain advice on fit, including when a shared Project is enough and when an API-built agent is the right build
- Month-to-month maintenance under managed AI services, from one office in Brooklyn, in business since 1998
Custom GPT, shared Project, workspace agent or API-built agent: which fits
OpenAI's own comparison calls a GPT single-player with static curated content and a shared project multi-player with content that evolves. The full picture, as of October 2026:
| Custom GPT (retiring) | Shared Project | Workspace agent or plugin | API-built agent | |
|---|---|---|---|---|
| What it is | A configured version of ChatGPT: instructions, knowledge, apps or Actions | A shared space of chats, files and instructions that a team keeps adding to | The newer in-ChatGPT builder: model, tools, apps, custom MCP servers, skills, files, schedules, triggers | Software built with the Agents SDK or Responses API that runs in your own systems |
| Best for | Reusable curated knowledge, such as an HR benefits assistant | Live team context that changes weekly, such as a planning hub | Recurring or scheduled work inside the workspace, with approvals on writes | Jobs that must run outside ChatGPT, embed in a website or act across several systems |
| Who builds and publishes | Members where settings allow; owners control sharing | Any member with the permission | Named people; publishing restricted by role | Developers, under an API project with its own identity and logs |
| Status as of October 2026 | Scheduled to retire December 11, 2026; migrate to a plugin | Current on all plans | Current on Business and Enterprise | Current; Agent Builder is deprecated, the SDK and Responses API remain |
| Where NetSys covers it | This page (migration) | This page and AI training | This page | AI agents for business |
More on the choice: ChatGPT Projects vs custom GPTs for a business team · Custom GPT vs AI agent: what should your company build? · AI agent development with approvals and logs.
How a custom GPT or workspace agent build runs
The order is the same whether the result is a plugin, a workspace agent or a GPT migration:
- Brief: one job, its inputs, approved sources, actions, audience and pass criteria, written on one page with a named owner.
- Fit and plan check: GPT, plugin, shared Project, workspace agent or API-built agent, and what your workspace settings allow.
- Knowledge: approved documents collected, old versions removed, formats and limits checked.
- Build: instructions drafted, knowledge attached, apps or MCP tools connected with read access first.
- Evaluation: the test set run, failures fixed, prompt injection checks on documents and sources.
- Controls: sharing set to the smallest audience, permissions and ownership recorded, write actions on always ask.
- Pilot: a small group uses it on real work for a few weeks, with a way to report wrong answers.
- Launch and maintenance: the employee guide published, the item added to the register, the monthly test run scheduled.
Deliverables, owners and boundaries
Each deliverable has an owner on both sides and a boundary written into the scope:
| Deliverable | NetSys | Your company | Boundary |
|---|---|---|---|
| One-page brief | Writes it from discovery | Names the owner and approves the scope | One job per assistant; new jobs are new briefs |
| Knowledge set | Assembles and checks it | Approves the documents and keeps them current | Only approved documents; no live client data in knowledge files unless the policy allows it |
| Instructions and configuration | Drafts, tests and versions them | Reviews tone and refusals | Built in your workspace under your owners |
| Integrations | Connects apps, Actions or MCP tools | Grants consent and owns the credentials | Read by default; writes behind an approval |
| Test set and results | Builds and runs it | Agrees the expected answers | Results recorded; no accuracy guarantee |
| Employee guide and register entry | Writes them | Publishes them | Kept in your workspace or SharePoint |
Examples of assistants we build (illustrative)
Illustrative patterns written for this page, not client assistants. Each would start with its own brief and test set:
| Job | Knowledge and access | Where it stops |
|---|---|---|
| Policy and HR questions | The handbook, benefits documents and the holiday calendar | Hands anything personal or disputed to HR |
| Proposal first drafts | Approved past proposals, service descriptions and the pricing rules document | Prices come only from the approved document; a person reviews before sending |
| Vendor document comparison | Contracts or quotes uploaded for the session | Flags clauses for review; gives no legal advice |
| Weekly report assembly (workspace agent) | Read access to a reporting folder, on a schedule | Drafts the report; a manager approves before it is shared |
| Support reply drafting | Product documentation and approved response templates | Drafts only; every reply is sent by a person |
Worked examples: Custom GPT examples for sales, operations and internal support · How to create a custom GPT for your business.
What moves the quote for a custom GPT or agent
NetSys publishes no price list. Each build is quoted as a fixed project from the one-page brief, and these are the inputs that move it:
| Factor | Why it moves the price |
|---|---|
| Number of jobs | Each assistant is its own brief, build and test set |
| Knowledge state | Clean, current documents are quick; scattered or duplicated ones need cleanup first |
| Integrations | Read access to an app is simple; a custom MCP server with write actions is a small development project |
| Audience and controls | Enterprise role design and allowlists add configuration and review |
| Migration | Existing GPTs to migrate, especially those with Actions that must be rebuilt |
| Maintenance | Monthly test runs and updates under managed AI services |
On the costs: Custom GPT cost: build scope, integrations and maintenance.
Your workspace seats and any credits are billed to you by OpenAI.
Relying on a custom GPT that retires in December?
Send its name, what it knows and whether it uses Actions. We will tell you whether it migrates cleanly to a plugin or should be rebuilt as a workspace agent.
Custom GPT Development FAQs
What is a custom GPT?
A custom GPT is a version of ChatGPT configured for a specific purpose: instructions, conversation starters, knowledge files, capabilities such as web search or code interpreter, and either connected apps or Actions. In a Business or Enterprise workspace they can be shared by invitation, by workspace link, with the whole workspace or to the GPT Store, and builders cannot see the conversations people have with them. Personal plans (Free, Go, Plus and Pro) can no longer create or publish them. Source: OpenAI help center, What are GPTs and Creating and editing GPTs, read October 17, 2026.
Are custom GPTs going away?
Yes, on OpenAI's published schedule. As of October 2026, custom GPTs are scheduled to retire on December 11, 2026, with February 11, 2027 for Enterprise workspaces granted a deferral, and creation of new GPTs on Enterprise planned to end on October 26, 2026. OpenAI recommends migrating to plugins: knowledge files become plugin reference files, GPT Actions do not transfer, and the migrated plugin starts private. Existing conversations stay accessible. We treat every GPT build or migration request against that calendar. Source: OpenAI help center article on the GPT retirement and migration, read October 17, 2026.
What is the difference between a custom GPT, a workspace agent and an AI agent?
A custom GPT answers inside ChatGPT from curated knowledge, with apps or Actions for reach. A workspace agent is the newer builder in ChatGPT Business and Enterprise: it has a chosen model and reasoning effort, tools, apps, custom MCP servers, skills, files, schedules and triggers, with write actions set to always ask by default. An AI agent in the sense of our AI agents page is software built outside ChatGPT, on the OpenAI API, in Copilot Studio or on Claude, that runs under its own identity, logs every action and waits for approval before it changes anything. Pick by where the job runs and what it must change. Source: OpenAI help center article on workspace agents, read October 17, 2026.
Can a custom GPT read or update our company systems?
Read, yes, within limits: connected apps such as SharePoint or Google Drive give read access inside each user's existing permissions, and OpenAI-built apps are search-only as of October 2026. Writing needs either a GPT Action (an OpenAPI schema with an API key or OAuth, which does not survive the migration to plugins) or a custom MCP server, which on Business only owners and admins can create and publish, and which ChatGPT uses as a frozen snapshot until an admin refreshes it. We build writes behind an approval step and prefer the MCP route or a workspace agent for anything that must last. Sources: OpenAI help center articles on configuring Actions, custom MCP servers and connected apps, read October 17, 2026.
Who can see a custom GPT we build, and does it learn from our conversations?
You decide the audience: invite-only, anyone in the workspace with the link, the whole workspace, or the GPT Store, with permission levels for chatting, viewing settings and editing, and direct sharing to up to 100 people or groups. On Enterprise, role-based access controls who may create and share, and an action domain allowlist governs Actions. GPTs do not use saved memory or previous conversations, and Business and Enterprise GPT data is not used for training by default. Source: OpenAI help center, Sharing and publishing GPTs and What are GPTs, read October 17, 2026.
How much does custom GPT development cost?
Each assistant is quoted as a fixed project from its one-page brief, so you see the price before work begins. What moves it: the number of jobs, the state of the knowledge documents, whether the assistant must write into a system (a custom MCP server is a small development project), Enterprise role design, any existing GPTs to migrate and the monthly maintenance you want. NetSys publishes no price list. Your workspace seats and any credits are billed to you by OpenAI. Our cost post walks through the factors.
Do we own the custom GPT, and who maintains it?
It lives in your workspace, under your owners, with the instructions, knowledge and test set handed over in writing; NetSys holds no copy you cannot see. Maintenance is a monthly cycle under managed AI services: the test set is re-run, knowledge and instructions are updated when the business changes, integrations are re-tested after vendor changes, and assistants without an owner are retired. When a builder leaves, ownership transfers to a workspace owner, which the offboarding runbook covers.
When should we build an API agent instead of a custom GPT?
When the job must run outside ChatGPT: embedded in your website or intake form, triggered by events in your systems, acting across several applications, or governed by retention and identity rules a workspace cannot meet. GPTs cannot be embedded in a website, and a ChatGPT workspace and an API organization are separate systems with separate billing. An API-built agent gets its own identity, least-privilege credentials, approval queue and logs, and it is billed per token rather than per seat. Our AI agents page covers that build. Source: OpenAI help center, What are GPTs and What is ChatGPT Enterprise, read October 17, 2026.
Sources and technical references
- OpenAI help center: What are GPTs (definition, privacy, GPTs versus API assistants)
- OpenAI help center: creating and editing GPTs (who can build, knowledge files)
- OpenAI help center: the scheduled retirement of custom GPTs and migration to plugins
- OpenAI help center: sharing and publishing GPTs
- OpenAI help center: configuring Actions in GPTs
- OpenAI help center: GPT admin controls in Enterprise and Edu (role-based access, action domain allowlist, ownership transfer)
- OpenAI help center: workspace agents in ChatGPT Business and Enterprise
- OpenAI help center: connecting custom MCP servers in ChatGPT
- OpenAI help center: Projects in ChatGPT, including the GPT versus shared project comparison
- OpenAI developer documentation: building agents on the API (Agents SDK, Responses API, ChatKit)
- OpenAI developer documentation: Agent Builder deprecation
- OpenAI help center: What is ChatGPT Enterprise (a workspace and an API organization are separate systems)
Guides on this topic
- AI agents for business: consulting and development
- Custom AI application development
- How to create a custom GPT for your business
- Custom GPT examples for sales, operations and internal support
- ChatGPT Projects vs custom GPTs for a business team
- Custom GPT vs AI agent: what should your company build?
- Custom GPT Actions: connecting business systems safely
- Custom GPT cost: build scope, integrations and maintenance
- OpenAI and ChatGPT integration services
- Managed AI services: monthly maintenance of GPTs and agents
- ChatGPT consulting and business workspace setup
- Corporate AI and ChatGPT training for employees
- MCP server development for Claude
- AI agents vs chatbots: what each one does for a business
- Case study: custom agents for an 18-person organization
Related services
Build the assistant on tools that will still exist next year.
Tell us the job, the documents and any systems involved. We will say whether it belongs in a plugin, a workspace agent or an API-built agent, and scope a pilot with a test set and an owner.