How PacedLoop works
PacedLoop is a workflow layer that runs behind a Custom GPT. ChatGPT handles the conversation. PacedLoop holds the workflow: the step pointer, the captured variables, and the rules that decide what the GPT can ask next. Every run is saved to a dashboard you can review. This page explains how that works.
The architecture
Client › ChatGPT › PacedLoop › Structured output
- Client
- The person you serve. They open the Custom GPT link and start typing. Nothing to install. A free ChatGPT account is enough.
- ChatGPT
- The conversational interface. Handles tone, follow-up questions, document uploads, image generation, and voice.
- PacedLoop
- The workflow engine. Holds the workflow definition, the current step pointer, the captured variables, and the access rules. Decides which step the GPT can ask next.
- Structured output
- One labeled record per run. Stored in your dashboard. Available for review, export, or downstream automation.
The GPT cannot advance the workflow. PacedLoop cannot run a conversation. Each layer does one job. That separation is the point.
Why state lives outside the GPT
Most attempts to make ChatGPT follow a process push everything into the system prompt: the steps, the rules, the prior answers, the next question. As soon as the conversation runs longer than a few turns, the prompt gets crowded, the model starts skipping ahead, and the captured answers disappear into the chat scroll.
PacedLoop moves the workflow state out of the prompt entirely. On each turn, the GPT receives only the current step’s goal, context, and expected output. Past answers exist as structured variables in PacedLoop, not as accumulated prompt history. The model’s reasoning stays local. The prompt stays bounded. The workflow stays deterministic across every run.
How much you keep in the hidden system prompt versus each step is a design choice. See the workflow patterns for four common approaches, including how to gate a GPT behind a name and email before it does any work.
Six technical questions about how PacedLoop works
If you’re evaluating PacedLoop technically, these are the core questions.
How is state stored?
The step pointer, the defined variables, and the captured structured outputs persist in PacedLoop. ChatGPT receives only the current step's payload on each turn. Past answers exist as queryable fields, not as accumulated chat history.
How is progression controlled?
PacedLoop decides which step the user is on. The GPT cannot advance unilaterally and the client cannot skip ahead. Each step defines what counts as a valid response. When the response satisfies the step, PacedLoop releases the next one.
How are outputs structured?
Every step has a title and a captured output. The step title is the field label. The user's answer, collected through the ChatGPT conversation, becomes the value. One run produces one structured record, ready for review, export, or downstream automation.
How are workflows versioned?
When a run starts, PacedLoop copies the entire workflow definition into the run record. That snapshot executes for the lifecycle of the run. Editing a workflow has no effect on runs already in flight or already complete. Historical answers stay attached to the workflow shape they were captured against. To iterate on a workflow without reconnecting the GPT, copy the live workflow, archive the copy as your historical record, and edit the live workflow in place. First-class versioning (named versions, side-by-side A/B, in-place rollback) is on the roadmap.
How are runs audited?
Every run lives in your dashboard with the full step-by-step capture: which step the user reached, how they answered each one, when each step was completed, and which fields were extracted. Nothing important is buried in a chat transcript.
How does human intervention work?
The end user is the human in the loop, by design. Every step requires a human response before the next is released. Pause and resume work natively while a run is active: the conversation persists in the user's ChatGPT thread, the workflow state persists in PacedLoop, and resuming is just sending the next message. Once a run is complete or 72 hours idle, recorded steps can be revised through repair: the end user requests the change through their GPT, supplies a reason, and provides the revised answer.
What a captured run looks like
Every step you define becomes a field on the run record. The step’s title is the field label. The user’s response, collected through ChatGPT, is the value. Below is a representative shape of what gets saved when a four-step consulting diagnostic completes.
{
"id": "run_2026_05_29_b4d6e7",
"workflowId": "wf_diagnostic_intake",
"status": "complete",
"currentStepIndex": 3,
"capturedContact": {
"name": "Alex Morgan",
"email": "alex@northwave.io"
},
"workflowSnapshot": {
"title": "Consulting diagnostic intake",
"steps": [
{
"title": "Name, company, and role",
"status": "completed",
"output": "Alex Morgan, Northwave Logistics, VP of Operations"
},
{
"title": "Primary challenge to address",
"status": "completed",
"output": "Order fulfillment SLAs have been slipping for two quarters. Warehouse throughput is the bottleneck."
},
{
"title": "Budget range and target timeline",
"status": "completed",
"output": "$75k to $150k for a 90-day engagement starting in Q3."
},
{
"title": "What does success look like in 90 days",
"status": "completed",
"output": "On-time delivery rate back above 95% and a roadmap to keep it there without overtime."
}
]
},
"createdAt": "2026-05-29T13:54:00Z",
"updatedAt": "2026-05-29T14:38:22Z"
}The same record is what you see in the dashboard, what you can export, and what downstream automation reads from. There is no separate extraction step.
How PacedLoop differs from the alternatives
PacedLoop sits between two common approaches. Developer frameworks give engineers powerful primitives but require writing and maintaining code. Custom GPTs alone give a finished conversational interface with no workflow integrity or captured state. PacedLoop fills the gap between them. See the detailed comparison.
- LangGraph
- Branching agentic graphs for engineers. Powerful primitives, custom code, ongoing maintenance. Use when you have engineers and need multi-agent orchestration.
- Custom GPTs alone
- A finished conversational interface with no workflow integrity, no captured state, and no review surface. Use when you want open-ended chat with no structure.
- PacedLoop
- A no-code workflow layer that runs behind the Custom GPT your clients already use. Deterministic progression, structured outputs, dashboard review. Use when you want a guided, reviewable process behind ChatGPT without writing code.
Who this is for
PacedLoop is built for coaches, consultants, and experts running structured client engagements: intake, discovery, diagnostics, assessments, onboarding.
The same architecture works anywhere a guided AI workflow has to behave predictably: hiring screens, customer success check-ins, training and certification flows, internal operations. Anywhere “AI follows your process” is more important than “AI improvises.”
Get started
Share the PDF, questionnaire, or framework you already use. We map it into a PacedLoop workflow, build the Custom GPT, connect it to your dashboard, and hand you back a live link. Free during beta. Prefer to build it yourself? Use the Workflow Builder GPT.