Mission Control

Mission Control turns one creator command into a governed, traceable workflow across specialized AI roles, tool adapters, and accountable human decisions.

Comment

Inspiration

AI can generate individual pieces of work, but people still have to coordinate prompts, tools, failures, approvals, and handoffs themselves. The missing layer is not another chatbot. It is one place that knows the objective, owns the critical path, explains responsibility, remembers decisions, and keeps execution moving.

What it does

The judged demo begins with Create Day 32. CEO Brain routes accepted outputs through Research, Script, Devotional, Voice, Visual, Thumbnail, SEO, QC, and Upload Package. Every stage exposes its executor, routing reason, state, and inspectable artifact contract.

After QC, Mission Control pauses at 86%. Upload Package remains blocked, the dashboard says a human is needed, and the creator can see why. Only after Approve & continue does the system record the decision, clear the blocker, activate Upload Package, and reach 100% Ready for YouTube. It prepares a creator-controlled handoff; it does not publish externally.

How we built it

  • A source-controlled contract defines stages, executors, dependencies, progress weights, approval policy, and artifacts.
  • A deterministic engine completes one eligible stage at a time, pauses after QC, and resumes only after approval.
  • Mission health, confidence, blockers, human need, and next action are recalculated from the same state.
  • Browser persistence preserves the approved mission after refresh.
  • The responsive dashboard renders engine state directly, without a separate fake progress model.

We used Codex with GPT-5.6 to recover and inspect the repository, reason across the state model, implement and test the approval transition, verify the browser flow, repair the capture pipeline, generate the narrated demo, and prepare the submission materials.

Challenges

The hardest part was making the approval boundary real rather than decorative. Progress, health, blockers, decisions, persistence, and the next action all needed to agree before and after approval. We also kept the demo honest: deterministic adapters are presented as inspectable contracts, not falsely described as live external integrations.

Accomplishments

  • Nine stages complete before a visible creator checkpoint.
  • The workflow pauses at 86% with Upload Package blocked.
  • Approval is recorded in decisions and recent activity.
  • Upload Package completes only after approval.
  • The mission reaches 100% Ready for YouTube.
  • 12 deterministic tests pass.
  • Browser verification passes with persistence, zero console errors, and no mobile overflow.
  • A verified editorial demo closes with the Mission Control pattern for long-running missions.

What we learned

Reliable agentic software needs visible ownership, durable decisions, explainable routing, and an explicit human-control boundary. Autonomy is valuable, but governed continuity is what lets real work finish safely.

What's next

Add the OpenAI provider behind the tested contract, then connect production voice, image, and YouTube adapters without changing the governance rules. Expand the mission schema so the same kernel can govern additional long-running workflows.

Current demo

Public demo video: https://youtu.be/2xNBpqFe0ro

Links

Built With

Share this project:

Updates