CoolFace
Apppublic

corrigibility/agent-oversight-architect

sourceHugging Facemitupdated 8d agoView on Hugging Face
0likes
App README

๐Ÿงญ Agent Oversight Architect

A browser-based control-plane designer by Corrigibility.

Design an agent workflow step by step and identify where execution should proceed automatically, pause at a checkpoint, require explicit human approval, or remain blocked until permissions are re-scoped.

What it designs

  • โ€”Approval gates
  • โ€”Stop points
  • โ€”Checkpoints
  • โ€”Permission scopes
  • โ€”Rollback requirements
  • โ€”Human escalation
  • โ€”Event logging
  • โ€”Retry limits
  • โ€”Post-intervention resume rules

Workflow risk factors

Each action can be configured by:

  • โ€”External impact
  • โ€”Reversibility
  • โ€”Permission level
  • โ€”Data sensitivity
  • โ€”Financial impact
  • โ€”Autonomy level

Outputs

  • โ€”Oversight architecture score
  • โ€”Per-action control classification
  • โ€”Human approval map
  • โ€”Stop / checkpoint plan
  • โ€”Rollback requirements
  • โ€”Permission policy
  • โ€”Machine-readable control-plane JSON
  • โ€”Markdown architecture export

Important

This Space is a design aid, not a security certification. Its recommendations are deterministic heuristics and should be validated against the actual system, threat model, policies, and operational environment.

Privacy

Everything runs locally in the browser. No workflow data is sent to a backend.

Tech

  • โ€”Static HTML
  • โ€”CSS
  • โ€”Vanilla JavaScript
  • โ€”No backend
  • โ€”No API
  • โ€”No external framework

Built for the Corrigibility organization on Hugging Face.