ActFlow

Inspiration

A task is not finished just because an AI has produced an answer or a tool has returned success. Real work has dependencies: documents need to be prepared, expenses need review, and the owner needs evidence that the intended result was actually delivered. ActFlow explores how an agent can carry a task forward while keeping those decisions visible and under human control.

What it does

ActFlow turns a goal and its supporting context into a persistent workflow. Its training-administration example prepares a participant spreadsheet and a response draft, pauses for an expense approval, records decisions, checks generated files, and asks the owner to confirm completion. The application keeps a task timeline and supports reviewing results after execution.

The original application includes policy checks, approval gates, artifact verification, task memory, event handling, and recovery for supported file changes. It distinguishes simulated actions from real external effects. For example, approving an expense in the demonstration records the approval; it does not make a payment. The demonstration uses synthetic employee data and a simulated external action.

The hosted English-language workspace implements a focused training-document workflow: create a task, enter or load sample data, approve or reject the expense review, generate and download real XLSX and DOCX files, inspect a verification summary, and confirm completion. A ChatGPT sign-in is required. The hosted version has a narrower scope than the original repository: Gmail and unattended scheduled processing have not been ported.

NVIDIA and Nebius integration

ActFlow includes an integration for NVIDIA Nemotron models through the Nebius Token Factory API, including the candidate model nvidia/nemotron-3-super-120b-a12b. In the hosted workspace, model-assisted source analysis produces suggestions for a person to review; it does not send email or execute external actions on its own. The submitted recording uses fixture-based execution with sample data and English display labels. Live Nebius inference is not established by the video or the offline tests.

How we built it

The original application uses Next.js, React, TypeScript, SQLite, ExcelJS, and DOCX generation. A worker and persisted events support task continuity on a running server. The hosted workspace uses a Cloudflare Worker with D1 for task records and R2 for generated files. Generated files are hashed and their stored bytes are checked when downloaded. Source code, the MIT license, setup instructions, and scope notes are in the public repository.

Challenges and what we learned

The hardest problem was defining completion. A plausible-looking file can still contain a wrong row, amount, recipient, or formula. In an offline synthetic evaluation dated October 23, 2026, the original verifier accepted 8/8 valid artifacts, rejected 12/12 deliberately invalid artifacts, and passed 4/4 workflow lifecycle cases. These 24 cases were run once each and involved no model calls or external sends; they are not a measure of real-world agent accuracy.

We also learned to treat approval, verification, and recovery as separate steps. A task should stay open when a result is missing, and an external effect should not be described as delivered merely because a tool was invoked.

What's next

Validate real NVIDIA inference through Nebius with a reproducible receipt, extend the hosted version with the original app's integrations and background processing, and evaluate the workflow with real users.

Built With

• cloudflare
• nebius
• next.js
• nvidia
• react
• sqlite
• typescript