Skip to main content
Use this workflow when an API client or external AI agent builds or extends a Vida Agent. The Vida skill contains the complete build and operation guidance. Apply only the steps needed for the requested capability and launch state. For Computer work, use Task → Skill → Helpers. The Task supplies run inputs and the requested outcome, the owning skill defines the procedure and recovery, and registered helpers perform the Computer operations. Ordinary voice and platform functions continue to use their documented APIs.

Required inputs

  • an authorized Vida API token
  • the organization account ID
  • the intended Agent account, or approval to create one
  • the business workflow and acceptance criteria
  • required access, authoritative state, output destination/retention, and per-run limits
  • approval for writes that can affect billing or production behavior
Keep targetAccountId, Agent-account IDs, and agentConfigId values distinct. Read IDs from Vida; do not reconstruct them from names.

1. Resolve the account

  1. Read the authenticated or selected account.
  2. Confirm its type, parent organization, enabled capabilities, and intended customer.
  3. List Agent accounts under the organization.
  4. Reuse the intended Agent account, or create one only after confirming the account and billing impact.
Establish the intended parent organization before creation. The token’s default organization proves access, not ownership intent. Infer already stated decisions rather than requesting them again. Save the returned Agent account ID. Use it as targetAccountId for Agent-scoped operations.

2. Discover available capabilities

Do not copy model, voice, function, or app identifiers from another account. Read the detail for every app or function that has setup requirements.

3. Read and update staging

  1. Read staging. Compare with live to identify unrelated unpublished changes before editing.
  2. Preserve writable values the user did not ask to change, especially authored arrays.
  3. Update staging with only approved writable settings.
  4. Read staging again and compare the complete persisted result with the intended configuration.
Agent instructions establish the role, boundaries, and routing. A Computer workflow’s detailed procedure and helper sequence belong in its owning skill, rather than a second copy in the prompt. Do not send calculated response fields, effective inherited values, or IDs from the live configuration back as authored staging settings.

4. Resolve dependencies

Before testing:
  • install and verify required apps
  • configure functions and their destinations
  • connect phone, SMS, SIP, email, or web channels as needed
  • provision and verify a Computer Agent before adding Computer-only capabilities
  • create reporting fields with stable keys
If a connection requires credentials, use its supported setup flow rather than adding credentials to Agent instructions. For Computer workflows, complete the owning skill and registered helpers. Keep required resources and retained file outputs with that skill. Declare authentication, callbacks, state, retention, and delivery dependencies. Prove runtime skill discovery separately from helper registration, then test representative helper execution and the destination effect. Use Browser exploration or recording only when needed to establish an operation. API and file helpers do not require a Browser recording. New workflows should not depend on heartbeats, conversational memory, workspace searching, or loose scripts.

5. Test the prepared Agent

Test the successful path, missing information, refusal or unsupported work, escalation, and every external side effect. Use a real representative channel when possible. Run the intended Task or request in a fresh execution without the builder’s prior conversation or temporary files. A caller that only needs one known helper can invoke it directly without creating an unnecessary Task. Preserve declared source-system state and authenticated Browser profiles. If recurring work is requested, create a repeating Task paused with instructions naming the owning skill and supplying its business inputs, scope, limits, and requested result. Inspect one controlled run and verify a second run skips or reconciles prior work. Activate only within the user’s authorization. See Tasks and automation. For an authorized outbound test, create a Task for the selected Agent and inspect its terminal result, attempt history, log, and conversation. Verify any external action in the destination system. Optionally review the Agent configuration, but do not treat automated review as a substitute for a representative test.

6. Save and publish

  1. Optionally save a named version for recovery or an experiment.
  2. Confirm the authorized release includes every staging/live difference. Publishing includes unrelated staged work too; preserve it and resolve release scope before proceeding.
  3. Publish staging.
  4. Read the live Agent independently.

7. Verify production

Run one production-path interaction and confirm:
  • the request reached the intended Agent and channel
  • the live configuration matches the approved staging result
  • functions, apps, and transfers worked in their destination systems
  • the log and complete conversation are available
  • reporting fields contain defensible values

Completion evidence

Report the organization and Agent account IDs, staging and live agentConfigId values, saved version ID when created, actual launch state, representative Task or conversation reference, and external verification. For Computer work, include the owning skill, registered helpers, state/output locations, and schedule state when applicable. Identify any unresolved dependency. Do not report completion from the publish response alone, or imply a staged delivery is live.