Aegis

Problem

AI agents operate with unchecked access to tools, APIs, and data. Without a control plane, there's no way to enforce what an agent can or cannot do, no tamper-proof record of what it did, and no way to reproduce a run for debugging or compliance review.

Approach

1. 1Capability manifests declare per-skill permissions (tools, network, budgets)
2. 2Policy engine evaluates every tool call against live session state
3. 3Cryptographically-linked audit log records every agent action
4. 4Deterministic replay rebuilds any session from the audit chain
5. 5Docker-based deployment with systemd support for production

System Pipeline

Evaluation

Aegis enforces least-privilege policies at runtime — any undeclared tool call is denied. The cryptographic audit chain provides tamper-proof compliance records, and deterministic replay enables exact reproduction of any agent session for debugging.

Learnings

• Capability-based security maps naturally to agentic tool-use patterns
• Cryptographic chaining catches subtle state corruption that simple logging misses
• Deterministic replay requires careful handling of non-deterministic LLM outputs