How does Miiwa turn a workflow into a reliable AI agent?
A Miiwa Agent is the controlled delivery form of a published native workflow: a validated graph, workspace-scoped services, persisted run state and a guided interface around the result.
In Miiwa, an agent is not a prompt with a broad tool attached. It begins as a native workflow in Studio and becomes a Miiwa Agent only when a validated version is published and delivered through a deployment. Reliability comes from that entire path.
Start with a native graph, not a single prompt
Native Studio represents a workflow as blocks and edges. Blocks collect input, scrape or retrieve context, call models, transform values, branch, use integrations and render outputs. Variables carry state between those blocks, so the contract is visible in the graph instead of being buried in one conversation.
User Input → Scrape URL → Generate Text → Generate AssetMiiwa Architect can scaffold the block combination, prompts and interface starting point from a business request. The builder then owns the concrete graph, variable names, service choices, tests and published version.
What does Miiwa check before publish?
| Boundary | Publish-readiness behaviour |
|---|---|
| Graph | The workflow must pass native graph validation rather than relying on a visually plausible canvas. |
| Blocks | Only block types on the production-ready whitelist can be published. |
| Human input | A User Input block needs fields and, when selected, a valid custom or canvas interface. |
| Asset delivery | A Generate Asset block must have the template or canvas required by its render mode. |
| Resources | Standalone resource definitions cannot be wired into execution edges. |
| Parallel work | Miiwa rejects pause/end blocks inside parallel branches and conflicting writes to the same output variable. |
How does a run survive human input?
When the engine reaches a User Input block, it emits a pause, records the current block and persists the run state. The user supplies the required values later. Miiwa then claims the paused session, merges the persisted variables with the new input and resumes from the known point in the graph.
This is why a guided Miiwa Agent is different from a stateless form wrapped around an LLM call. The workflow can stop, wait across requests and continue without rerunning the earlier blocks or asking the user to reconstruct hidden state.
One runtime serves every supported trigger
| Trigger path | How it enters Miiwa |
|---|---|
| Assigned Miiwa Agent | A user prepares guided input and starts an on-demand run from the dashboard. |
| Webhook | An authenticated deployment endpoint creates work from an external event. |
| Inbound email | A configured deployment accepts email as trigger context. |
| Browser extension | Selected browser context becomes input to a deployed workflow. |
| Schedule | Miiwa creates queued work from a configured automation schedule. |
| MCP | An assigned deployment is exposed as a tool through the workspace MCP surface. |
Where do models and credentials fit?
Native AI blocks use Miiwa’s model abstraction and Service Router instead of reading provider keys directly. Credentials can be resolved for the relevant user, workspace or platform context without appearing in the workflow or end-user interface. Required integrations are checked at the run boundary so an assigned agent does not quietly start without a connection it depends on.
What evidence does a run leave?
- A persisted run session with deployment, version, workflow, trigger and workspace context.
- Block-level events for starts, outputs, pauses and errors.
- Variables and checkpoints needed to resume the workflow safely.
- A final output payload, rendered asset or other delivery artifact when the run succeeds.
- Duration, error and cost context for operational inspection where the corresponding service records it.
A Miiwa release loop
- Use Architect or Native Studio to create the first graph and interface.
- Name the input and output variables so every downstream dependency is explicit.
- Run the workflow in Studio with representative and failure-inducing inputs.
- Resolve publish-readiness errors and review warnings for the chosen triggers.
- Publish the validated version and assign it to a narrow workspace or user group.
- Inspect run history, pauses, errors, outputs and costs before expanding the assignment or triggers.
A reliable Miiwa Agent is not a more persuasive prompt. It is a validated workflow version running inside Miiwa’s workspace, state, service and delivery boundaries.
Common questions
- Can a better system prompt replace Miiwa’s workflow controls?
- No. A prompt can guide model behaviour, but it cannot replace graph validation, workspace authorization, credential resolution, persisted pause/resume state, assignment rules or deterministic variable handling.
- What happens when a Miiwa Agent needs more information from the user?
- A User Input block pauses the native run and persists its state. After the user submits the required fields, Miiwa resumes the same run with the new values bound into its variables.
- Do scheduled, webhook and dashboard runs use different workflow engines?
- No. They are different entry paths into the native runtime. Each path must establish the correct workspace, deployment, trigger and visibility context before the workflow executes.