The loop
From an idea to production — on a loop you control.
Describe a process and Dooap Shepherd builds it. Prove it in UAT, gate every decision, promote what’s validated, then watch each run and improve on purpose. The cycle tightens over time.
Step 1 · Describe
Describe the process. It designs the path.
Open Dooap Shepherd beside your work and say what you want to happen. It designs the Action from what actually exists in your tenant — your step types, models, and tools — and shows you a preview before anything is created.
Meet Dooap ShepherdDooap Shepherd
press S anywhere
When an invoice arrives from a brand-new vendor, check the bank details against our history and hold anything suspicious. Show me a preview first.
Dooap Shepherd
✓ step types ✓ tenant tools ✓ design validated
I’ll extend Invoice intake with three steps after vendor identification:
AGENT STEP
Screen bank details
gpt-5-mini · tools: Search history
SWITCH STEP
Suspicious?
routes: hold · continue
TRIGGER ACTION
Hold & notify AP
async · tracked with this run
Preview ready. Nothing is created until you approve.
Step 2 · Build
An Action takes shape — typed, named steps.
What you get is a workflow you can read: OCR, agents, code, switches, tool calls. Each step does one thing; outputs flow forward, and branching and errors get their own dedicated types. Everything Shepherd would create or change is a preview you approve.
Explore the platformStep 3 · Prove in UAT
Prove it where it can’t do damage.
The fast, generative part lives in a UAT tenant that mirrors production. Dry-run on real-shaped data and evaluate against datasets — generative speed, nowhere near the real ledger.
Evaluation & datasetsStep 4 · Gate
Every decision passes a gate.
Confidence gates sit on every agent. Above the line, the work posts automatically; below it, the Action routes to a human instead of guessing. The same trust layer, on every run.
- Read document (OCR step)
- Extract fields (Agent step)
- Validate (Code step)
- Route by confidence (Switch step)
- Send downstream (Tool Call)
- High confidence: posted to the ERP. Below the gate: routed to a human.
Step 5 · Promote
Promote the version you proved.
Export the finished Action and its Apps and import them into production. What goes live is exactly the version you validated — and it arrives disabled, with credentials and secrets staying put. Run it once, check the result, enable it.
How promotion worksBuild & validate
Auto-build all you want.
describe → auto-build → dry-run → evaluate
Runs the real work
Only what you promoted.
gated · logged · reversible
✓ Credentials and secrets never leave production.
Step 6 · Observe & improve
Watch every run. Improve on purpose.
Every run is logged with its reasons, tool calls, and results — inspectable end to end. Promote interesting runs into test cases so they’re checked forever, then re-run the datasets to confirm a change is an improvement, not a regression. That’s the loop — each version better than the last.
Observability & evaluationBring this loop to your team.
We’ll show how Dooap Studio shortens the cycle from a plain-language idea to a gated production Action.