There's a subtle but critical gap in how most systems handle access control: you know exactly when your database recorded the revocation, but you've got no idea when that change actually propagates to every active session, cached token, or downstream service that depends on it. NexuChat ran straight into this problem maintaining a web application with multiple roles and language configurations, and decided to build Parallax as their entry for the All Things Agentic Hackathon.
The Revocation Timing Problem
Traditional access control assumes revocation is instantaneous. You flip a flag, update a role table, or invalidate a tokenβand you're done. Except you aren't. Active sessions still hold stale permissions. API keys cached in memory keep working until the next restart. Third-party integrations keep hammering your endpoints with credentials that are technically dead but functionally alive. The timestamp shows "revoked," but the access hasn't actually stopped.
Why Agentic Systems Amplify This
AI agents and autonomous workflows make this problem worse, not better. When you hand an agent a long-lived API key or OAuth token, you're creating a window where that credential can be revoked on your side while the agent happily continues using it on theirs. The agent doesn't know the rules changed untilβwell, often never, unless you've built explicit revocation checking into its runtime behavior.
Parallax: Tracking the Gap
The Parallax project (available on GitHub at github.com/NexuChat/parallax) tackles this by creating a system that tracks not just when access was revoked, but when it actually stopped being exercised. This requires correlating revocation timestamps with active session monitoring, token usage patterns, and real-time permission checks.
What This Means for Developers
If you're building anything with multi-user roles, API integrations, or AI agent workflows, you need to think about revocation as a process, not an event. That means instrumenting your systems to detect when previously-valid credentials are still being used after their official death dateβand alerting on it.
Key Takeaways
- Access revocation timestamps and actual access stop times rarely align in production systems
- Agentic workflows with long-lived credentials create particularly risky windows for this gap
- Monitoring "revoked but still active" patterns is just as important as the revocation itself
- The All Things Agentic Hackathon produced tools addressing these real-world gaps
The Bottom Line
Until your monitoring tells you when access stopped, not just when you revoked it, you've only solved half the problem. Parallax and similar projects are pointing at the other half that most of us have been ignoring.