## Summary
Brings the workflow AI/MCP tooling for **AI Agent steps** to parity with
the existing **CODE / logic-function** flow, and makes AI Agent output
references reliable in validation and the variable picker.
Just like a CODE step needs a logic function, an `AI_AGENT` step needs
an agent. The agent is already created as a side effect of step
creation; this PR adds the missing "configure it" tooling and fixes the
output schema so it reflects the agent's actual response format.
## Changes
### New `update_agent` MCP tool
- `update-agent.tool.ts`: lets the assistant configure the agent backing
an `AI_AGENT` step — `prompt` (system prompt), optional `modelId`,
optional `responseFormat` (text or structured json) — via
`AgentService.updateOneAgent`. Direct analog of
`update_logic_function_source`.
- Wired through: added `agentService` to `WorkflowToolDependencies`,
injected `AgentService` and registered the tool in
`workflow-tool.workspace-service.ts`, imported `AiAgentModule` in
`workflow-tools.module.ts`.
### Guided creation flow (mirrors CODE)
- `create-workflow-version-step.tool.ts`: `enrichResultWithNextStep` now
returns an `AI_AGENT` hint instructing the assistant to call
`update_agent` with the step's `settings.input.agentId` (and to set the
task prompt via `update_workflow_version_step` if needed).
- `create-complete-workflow.tool.ts`: rejects `AI_AGENT` steps (it
inserts steps directly and never runs the side effect that creates the
agent), with a description note pointing to
`create_workflow_version_step` + `update_agent`. Same treatment CODE
already gets.
### Correct output schema for AI Agent steps
- Backend `computeStepOutputSchema`
(`workflow-schema.workspace-service.ts`): the `AI_AGENT` case now
derives the output schema from the agent's `responseFormat` instead of a
hardcoded `{ response }`:
- text → `{ response: string }`
- json → one leaf per `responseFormat.schema.properties` field
This makes workflow validation resolve `{{stepId.fieldName}}` references
against the agent's real output (previously json agents validated wrong:
real fields rejected, `{{stepId.response}}` accepted but undefined at
runtime).
- Frontend `useStepsOutputSchema.ts`: when an `AI_AGENT` step has no
persisted `outputSchema`, fall back to generating it from the agent's
`responseFormat` (via `FindManyAgents` + the existing
`agentResponseSchemaToOutputSchema`) instead of the hardcoded `{
response }`. Keeps the variable picker correct for structured agents.
## Why output schema matters
Validation resolves every `{{stepId.path}}` against the referenced
step's `settings.outputSchema` (`validateWorkflowVariableReferences`).
The agent step's output schema is therefore the single source of truth
for "is the right output referenced." Because the runtime output depends
on `responseFormat` (text → `{ response }`, json → schema fields
directly), the schema must be derived from `responseFormat` to be
accurate.
## Not solved issue
We want each AI_AGENT step's settings.outputSchema to always match the
backing agent's responseFormat:
- text → { response }
- json → one field per responseFormat.schema.properties
That output schema is what everything downstream relies on: validation
(validateWorkflowVariableReferences resolves {{stepId.field}} against
it), the frontend variable picker (useStepsOutputSchema), and it's also
persisted inside workflowVersion.steps.
Agent should be unique source of truth but syncing agent -> step is not
possible
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/21755?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"></picture></a>
<!-- End of auto-generated description by cubic. -->
Etienne
d67aa2889b
