Coming back
Having been through the programme as a mentee, I came back to run it — mentoring Ashutosh and MHD. Ali Haider over six months on secure infrastructure for autonomous AI agents.
Caracal — authority delegation for AI systems
- Problem
- Humans, agents and services constantly delegate authority to each other, with nothing checking policy before execution, no way to revoke access in real time, and no record you could audit afterwards.
- Approach
- Led the architecture and guided the implementation: a policy enforcement gateway in front of every request, a delegation graph engine with authority evaluation over it, vault integration for API credentials, immutable audit infrastructure, and end-to-end delegation workflows tying them together.
- Result
- A cohesive, production-oriented foundation rather than a collection of independent features.
What the contributors took away
Graph-based authority resolution, policy-driven request routing, secure credential management, decentralised identity, and ledger-backed auditability for autonomous systems.
Most of them had not touched security infrastructure before. By the end they were making design calls in it.
Running the mentorship
Breaking the roadmap into milestones, reviewing pull requests, holding design discussions, documenting the architecture, and working through the problems that do not have a clean answer.
The part that mattered most was explaining the reasoning behind each decision, not just the implementation. A contributor who knows why can make the next call without me.
Review is teaching. If a mentee leaves a PR having only learned that they were wrong, I wasted both our time.
Taking it out into the open
The project was presented at several open source events and summits, giving visibility to the work happening inside LF Decentralized Trust and showing a practical approach to secure authority delegation for AI systems.