Open Source in Finance Forum London 2026

Multi-agent systems in regulated environments introduce a new class of operational risk: not whether agents can act, but how far their authority extends and what stops them from exceeding it. Autonomy creep is a hard problem, especially when agents can spawn and orchestrate other agents. This session presents the Autonomy Levels (A0–A4) defined in the AIGF Multi-Agent Reference Architecture (also merged into the AIGF taxonomy) and a pragmatic, cloud native way to enforce those guardrails from low to high autonomy. The approach targets real-world stacks built on Kubernetes and Flux CD, using a simple control model: agents propose, humans approve (where required), Flux reconciles, and the cluster enforces policy where admission control can make autonomy policies enforceable at runtime by consulting external sources of truth, e.g. an agent registry / entitlement list (is this agent allowed to run?). Unregistered agents, and unapproved “spawn-a-new-agent” attempts are rejected before they land. A service mesh policy layer limits A2A communication, and coordination stays within approved boundaries. Stack logs are consumed in real-time to understand policy-violating actions.