Skip to main content
Completed work gives your team evidence for the next case: what fixed the problem, which questions were missing, where the agent got stuck, and whether the customer received the right result. Studio can turn eligible evidence into knowledge and improvement proposals. Your team controls what is shared and which workflow changes take effect. Learning has three distinct parts: remember useful facts, reuse a proven method, and improve the agent’s workflow. They have different permissions and review rules.

​
Start with an outcome you can trust

Open the agent’s Outcomes view. Review the result alongside the task’s Activity, outputs, human decisions, and live-system evidence. Confirm or correct the outcome judgment where you have the authority to do so. Compare similar cases across versions. Success rate, cost per completed item, time, and repeated escalations can show where to investigate. A score or a cheaper run alone does not prove that the customer received a correct resolution.

​
Keep the useful knowledge

With Org Brain enabled, eligible sources are indexed with their visibility preserved. Completed and failed agent reports can provide channel-scoped evidence. The scheduled Knowledge writer can extract durable facts with links back to their sources. An 85% model-confidence threshold is one requirement for automatic fact publishing, not a guarantee of correctness or permission to widen its audience. Review source evidence, retract inaccurate facts, and correct the underlying source when needed. Workers currently search organization-visible evidence only. A useful channel lesson does not automatically become searchable by cloud agents. If the lesson is suitable for the whole organization, save a reviewed organization-wide Memory with only the information that audience should receive. See Org Brain.

​
Reuse a successful method

Use Save as procedure on a suitable conversation or completed procedure-run report to create a draft runbook. Review its arguments, tools, steps, and success criteria before saving it. A procedure captures the method, such as checking a service, performing a bounded recovery, and verifying health. An agent can use an eligible saved procedure as a workflow step with defined inputs and a saved snapshot. The cloud step still needs supported cloud tools and the agent’s connections; it does not inherit desktop access. Desktop procedures have their own refinement behavior. Eligible low- or moderate-risk changes can be applied automatically when auto-refinement is enabled; other proposals need review. This is separate from cloud-agent workflow refinement. See Procedures for controls and limits.

​
Improve the agent’s workflow

The worker-refinement process uses recent outcomes to propose changes to the agent’s stage instructions. In Outcomes, inspect the proposal, rationale, and available evaluation evidence before applying it. An organization admin can apply a supported workflow change to create a new worker version. Advice-only proposals are acknowledged without changing the workflow. Revert restores the earlier stages as another version. Tasks already in progress keep their pinned version. Current worker refinement is proposal-based: it does not silently rewrite the active workflow. After changing a workflow, run representative cases again. Include an exception, a rejected action, and a failed verification. Look for a better support result, not just fewer questions or lower cost.

​
Keep the review loop practical

  1. Review unresolved cases and incorrect outcomes first.
  2. Correct a mistaken fact in its source and retract an incorrect derived fact.
  3. Capture a proven operating method as a procedure when it is repeatable.
  4. Review and apply a workflow proposal only when the change addresses a real pattern.
  5. Test the next version and compare its results with earlier work.
Configure the knowledge jobs and inspect skipped or failed runs under Customize → Org Brain → Dream settings and Dream runs. Worker improvement proposals are managed separately in the agent’s Outcomes view. Model policy, budgets, source eligibility, and available evidence affect whether either process can run. Learning never grants permission to change a system. Tool policy, credential scope, required human review, and explicit endpoint exceptions still govern the next action.