Ideas · Builder Notes
Agent-to-agent delegation and authority lineage
Kognita alpha 0.5 lets agents delegate only narrower authority, and kognita evidence authority traces any action back to its human source.
What was wrong?
Multi-agent systems blur who authorised what.
What did I build?
Run.delegate_to passes attenuated, narrow-only scope, and it is evidenced. kognita evidence authority traces an action's lineage. Agent-to-agent calls go through the gateway path with classifier envelopes and taint. External agents authenticate at the lowest trust.
Per-agent quotas, run isolation and sandbox constraints for code execution were added and evidenced.
What surprised me?
Authority can only shrink as it passes down a chain.
What's still open?
Replaying child delegations needed a fix later.
What would I tell another builder?
Make delegation narrow-only by construction. Convention will not hold.
Sources
Kognita is an open-source alpha. This note comes from Kognita PR #60 in the Kognita repository.