Have a project in mind? Let's talk.
OpenAI AgentKit sunset planning is now a production task, not a product note. OpenAI says Agent Builder and Evals are winding down and will no longer be available from November 30, 2026. ChatKit remains available. OpenAI points code-first workflows to the Agents SDK and natural-language team workflows to Workspace Agents in ChatGPT. Source
This checklist is for teams that built useful Agent Builder workflows and now need a safe path off them. It is our synthesis of current OpenAI migration, SDK, Workspace Agents and MCP documentation, not a claim about a Codesprint-managed migration.
The deprecation is narrower than the whole AgentKit label. OpenAI's Agent Builder page says the visual workflow product is scheduled to shut down on November 30, 2026, while ChatKit remains available. The separate Evals product is also being wound down. Source
Our first step is an inventory split by dependency, not by team memory. Record every Agent Builder workflow, owner, business use, ChatKit embed, connector, MCP server, file source, scheduled trigger, approval step, eval dataset and production user group.
Mark each item as migrate, replace, retire or unknown. Unknown workflows need an owner and a deadline. A forgotten workflow can fail quietly even when the main customer-facing agent has already moved. Include prototypes that still receive real traffic, since they carry the same shutdown risk.
OpenAI's migration guide lets you open a workflow, select Code, choose Agents SDK and copy a TypeScript or Python export. It also warns that the export does not convert the workflow graph and does not guarantee that every behavior transfers unchanged. Source
Our rule is to treat the export as source material for a reviewed rebuild. Use it to recover prompts, model settings, tool wiring and intended flow. Do not paste it into production and assume parity.
Create a behavior baseline before changing anything. Save representative inputs, expected outputs, tool side effects, refusal cases, approval cases, latency, cost and failure handling. These become the acceptance tests for the new implementation.
OpenAI positions the Agents SDK for building agents through code. The SDK includes agents with instructions and tools, handoffs, guardrails, sessions, human-in-the-loop mechanisms and built-in tracing for debugging and monitoring workflows. Source
Choose this path when the workflow needs version control, code review, CI, deterministic control flow, custom deployment, provider-specific infrastructure or detailed observability. It is also the safer default when a wrong tool call has financial, security or customer impact.
Rebuild the workflow as explicit application behavior. Keep prompts, policies and tool schemas in reviewable files. Add tests for routing, tool arguments, approvals, retries, timeouts and final output shape. Use the SDK's tracing plus the production signals in our AI agent observability checklist.
OpenAI positions ChatGPT Workspace Agents for natural-language workflows shared with teams. Its help center says Workspace Agents can be tested before publishing, connected to apps and tools, shared with teammates, used in Slack, scheduled and triggered through an API. Source
Choose this path when the workflow is a repeatable team process and the owner does not need a custom application runtime. Candidate examples include meeting follow-up, internal research, procurement routing, support triage and recurring reporting.
Do not force highly deterministic workflows into this path. OpenAI's migration guide warns that workflows with strong determinism at their core may not migrate faithfully to a Workspace Agent. Source
Workspace Agents are off by default at launch for ChatGPT Enterprise workspaces, and admins enable them for eligible workspaces. Role-based access controls apply: users can only see and run agents they have permission to access. Source
Our migration checklist recreates the access model explicitly. Record who can chat, who can edit, who owns the agent, who can publish, which groups receive access and whether the agent is private, link-shared or listed in the workspace directory.
Review connected account mode before launch. OpenAI distinguishes end-user account authentication from an agent-owned shared connection, recommends a service account where possible and advises limiting access to what the agent needs. Source
OpenAI's Agents SDK supports hosted MCP tools, Streamable HTTP servers, legacy SSE and stdio servers. Its guidance says MCP tools can expose data and act with provided credentials, so teams should trust servers before connecting, use least-privilege credentials, keep tokens out of URLs and require approval for sensitive operations. Source
Our connector review lists every MCP server, transport, credential owner, tool scope, approval policy, network path and data class. Prefer Streamable HTTP or stdio for new integrations; OpenAI notes that the MCP project deprecated SSE and recommends keeping it only for legacy servers. Source
For sensitive tools, require approval per tool rather than relying on a broad agent-level warning. The SDK supports approval policies for local MCP tools and hosted MCP tools. Pair the migration with our MCP server security checklist.
The Evals shutdown is easy to underestimate because the workflow may still run while quality checks disappear. Before migrating, export or recreate the datasets, graders, prompts, pass thresholds, trace review habits and release decisions that depend on Evals.
Our replacement flow uses three layers: pre-production regression tests, sampled production evaluation and scheduled red-team or adversarial checks. The destination can be SDK code, your existing CI system, a dedicated eval platform or a mix. The important part is that no deploy goes live without a quality gate.
Keep old and new behavior comparable. Run the same representative inputs against Agent Builder and the replacement, then score task completion, groundedness, tool-call accuracy, policy adherence, latency and cost.
Our practical cutover sequence is:
Freeze workflow edits except urgent fixes.
Export code and save prompts, versions, connector settings and screenshots of the graph.
Build the replacement in a separate environment.
Recreate credentials with least privilege; do not reuse broad tokens because they are convenient.
Run baseline inputs against old and new paths.
Test failure, timeout, approval, retry and no-permission paths.
Start with one internal team or low-risk workflow.
Keep the old workflow read-only during a short comparison window.
Switch traffic, monitor traces and roll back by routing users to the old path if needed.
Retire unused connectors, credentials, schedules and eval datasets after the new path is stable.
Do not wait for late November to discover that a connector, admin permission or Workspace Agent feature needs another team's approval.
Migration is also a chance to remove hidden coupling. Check hard-coded users, personal connector accounts, undocumented schedules, Slack channels, API triggers, shared files, prompt experiments and manually maintained datasets.
OpenAI says Workspace Agent API triggers queue a run and return 202 Accepted; access tokens are scoped to Workspace Agents API operations. Treat those tokens like production credentials and monitor queued runs like any other job system. Source
For SDK migrations, verify runtime configuration, tools, authentication, permissions and deployment before release. OpenAI explicitly lists those as validation requirements for an Agents SDK implementation. Source
Codesprint helps teams move agent workflows into reviewed code, governed Workspace Agents and production AgentOps processes. Contact us if the migration touches customer data or side-effecting tools.
No. OpenAI says Agent Builder and Evals are winding down, while ChatKit remains available. The recommended destinations are the Agents SDK for code workflows and Workspace Agents for natural-language team workflows. OpenAI AgentKit update
No. OpenAI says the export does not convert the workflow graph or guarantee unchanged behavior. Use it as the starting point for a reviewed rebuild and behavior comparison. OpenAI migration guide
Repeatable team processes that fit natural-language configuration and ChatGPT sharing are the best fit. Highly deterministic workflows or custom production applications usually belong in the Agents SDK. OpenAI Workspace Agents
Start with a useful but low-risk workflow that has a clear owner, limited tools and representative test inputs. Use it to prove the export, access, evaluation, monitoring and cutover process before migrating high-impact agents.
Ready to take the first step towards unlocking opportunities, realizing goals, and embracing innovation? We're here and eager to connect.