A useful AI workflow should survive a vendor decision. If the tool changes price, removes a feature, loses an integration, or shuts down a product, the business should still know what the workflow does, which sources govern it, what good output looks like, and how to move the work elsewhere. Otherwise the company did not build an operating system. It rented one without keeping the plans.

OpenAI's AgentKit announcement now carries a June 2026 update saying that Agent Builder and the hosted Evals product are winding down and will no longer be available after November 30, 2026. OpenAI recommends moving code-based workflows to the Agents SDK and natural-language use cases to Workspace Agents in ChatGPT. That is not evidence that OpenAI's replacement paths are weak, nor that visual builders are a mistake. It is a clean reminder that even first-party AI products can have short planning horizons compared with a contractor's estimating, purchasing, client-service, or closeout process.

Own the workflow before you choose the tool

The durable asset is not the canvas, prompt, or model name. It is the operating specification: the event that starts the job, required inputs, governing sources, state transitions, allowed actions, approvals, retry rules, stop conditions, output format, acceptance test, and system of record. Write those elements in plain language outside the vendor interface.

  • Trigger: the exact user action, schedule, email, upload, or system event that starts the work.
  • Sources: the contracts, drawings, price books, project records, policies, or approved templates the workflow may use.
  • States: received, validating, missing information, processing, awaiting approval, completed, failed, and archived.
  • Controls: permissions, financial thresholds, client-facing commitments, escalation rules, and named human owners.
  • Evidence: inputs, source versions, tool calls, outputs, user edits, approvals, errors, timestamps, and final disposition.
  • Acceptance: a small set of real examples and edge cases the replacement must pass before it handles live work.

Separate business logic from vendor configuration

Suppose an AI workflow compares subcontractor bids. The business logic says every bidder, base amount, alternate, allowance, exclusion, schedule note, insurance gap, and unresolved question must appear with a source reference. The vendor configuration expresses that rule as nodes, prompts, tools, or code. Keep the business logic in a readable versioned document so it can be rebuilt in another environment without reverse-engineering screenshots.

Do the same for source access. Record which system contains each fact, how identity and tenant boundaries are enforced, how freshness is checked, and which actions are read-only. Anthropic describes agent behavior as the interaction of four layers: model, harness, tools, and environment. Portability requires documenting all four. Moving the prompt while forgetting permissions, tool behavior, or runtime access does not reproduce the workflow.

Keep your evaluation cases portable

A hosted evaluation dashboard can be convenient, but the underlying test cases should belong to the business. Store sanitized inputs, expected fields, prohibited actions, source requirements, reviewer decisions, and failure labels in a format that can be exported and rerun. Keep confidential project data inside the approved environment; portability does not mean copying customer records into an uncontrolled spreadsheet.

Before a migration, run the old and replacement workflows against the same frozen test set. Compare missing-item rate, source accuracy, unauthorized-action rate, first-pass acceptance, reviewer minutes, latency, cost, and escalation quality. A successful migration is not one that produces similar prose. It preserves the operating outcome and its controls.

Design the record trail for a handoff

For every important run, preserve the workflow version, model, source identifiers, tool calls, intermediate status, final artifact, human edits, approval, and error history. Decide how long those records must remain available and how they can be exported. If a tool disappears, an auditable history lets the replacement team understand what actually happened rather than reconstructing the process from employee memory.

Use a quarterly portability check

Once a quarter, choose one production workflow and answer five questions: Can we export its instructions and test cases? Can we identify every source and permission? Can we reproduce one completed run? Can we disable it without losing business records? Could another qualified person rebuild the minimum safe version from the documentation? Any no becomes a concrete backlog item.

Publish durable operating knowledge

Google says there is no special AI-only schema required for AI Overviews or AI Mode. Helpful, original, technically accessible content and accurate structured data remain the foundation. The durable value here is not the product retirement headline. It is the exit-plan checklist a building company can use to protect its workflows, evidence, and customer obligations as AI platforms change.

Continue with the operating system

Sources Read

Next step, if this note maps to a problem on your desk: Private Training — a private working session for your team ($1,500+).

Related Field Notes