Notes on building governed, mission-critical AI agents: runtime, policy, memory and knowledge graphs.
Agents that act on real systems need the same authorization discipline as people. Here is how A3T puts a Cedar check in front of every step.
Read the postAgents forget between runs unless memory is designed in. A short tour of how A3T stores what agents learn and lets you query it with AQL.
Why A3T runs agents as CNCF Serverless Workflows on a Rust runtime, and what that buys you in latency, cost and auditability.
SecOps triage, trading checks, RTB bidding and integration agents. Where governed agents pay off fastest, and what each one needs.
Sample note: puts structured knowledge into the model’s attention instead of the prompt. Your commentary goes here: why it matters for A3T and what you’d do differently.
Sample note: the design behind the policy language A3T uses on every agent action. Add your take here.
Sample note: treats the context window like memory tiers. Compare with how A3T keeps agent memory outside the model.
Sample note: the open workflow format A3T agents are defined in.