In-worker Mastra runtime
Agents and workflows execute inside the dashboard API Worker.
Built inside Smicolon:Edge orchestration.An AI workflow we run in production every day.
A technical deep dive into durable Mastra workflows running inside a Cloudflare Worker with D1 state, queues, model routing, telemetry, and Payload publication.
- Built by Smicolon, for Smicolon
- Designed, shipped, and operated in-house.
- Human-governed
- A person approves anything that leaves the system.
- Observable execution
- Runs, state, and failure paths remain visible.
- In daily use
- Smicolon's own team works in this system.
Edge runtime
Durable state
D1 records workflow state and snapshots beyond a single request.
Queue boundaries
Consumers, retries, and dead-letter paths separate request handling from long-running work.
Traced execution
Every workflow step emits telemetry; Mastra and Langfuse turn it into a readable trace.
Inside the system
Each part is a working dashboard or worker surface in Smicolon's own stack.
In-worker Mastra runtime
Agents and workflows execute inside the dashboard API Worker.
Durable state
D1 records workflow state and snapshots beyond a single request.
Queue boundaries
Consumers, retries, and dead-letter paths separate request handling from long-running work.
Traced execution
Every workflow step emits telemetry; Mastra and Langfuse turn it into a readable trace.
How work moves through the system
Inputs, decisions, bounded automation, human authority, and state that survives a restart — the whole loop, not one clever step.
Enqueue
Persist the requested operation and hand execution to a queue boundary.
Route
Select the bounded workflow, tools, and model configuration for the task.
Persist
Record stage progress, outputs, errors, and reconciliation state in D1.
Integrate
Publish approved results into Payload or another explicit downstream surface.
Running today. Built to go further.
Durable AI workflows run at the edge, inside the Worker itself — state in D1, queues for the long-running work, traces on every step. The models come from managed routing today; because routing is a boundary in the architecture rather than a dependency, the same workflows move to a dedicated or private runtime the moment your data boundary requires it.
Running in production
Smicolon designed this workflow, ships changes to it, and operates it as part of daily delivery.
Set per engagement
Data boundaries, model routing, approval depth, and integrations are decided against your constraints — not inherited from ours.
Bring us the workflow that is still held together manually.
We will map the shortest safe path to a production workflow — through the systems, approvals, data boundaries, and exceptions you already live with.