Notes from
the deep.
Architecture decisions, governance patterns, and the occasional whitepaper — written engineer-to-engineer.
all whitepaperblognautilusgovernancebosunarchitecture
Two Independent Stop Layers: Why One Budget Cap Isn't Enough for Autonomous Agents
Why a single configurable budget cap is a single point of failure for self-modifying agents, and how auto-r-graph proves a second, untouchable stop layer with a test that disables the first.
Policy-First Data Access: Why Your AI Agents Need a Broker
Direct database access for AI agents is a liability. A data broker with three personas — Librarian, Curator, Gatekeeper — gives you deterministic control over what agents know.
Cumulative Exposure Tracking in Multi-Agent Systems
Individual data requests look harmless. Accumulated across agents and sessions, they build profiles that shouldn't exist. Cumulative exposure tracking makes the invisible visible.
Why Expert Systems Are the Missing Governance Layer
Policy engines evaluate rules. Expert systems reason about state. Here's why that distinction matters for AI agent governance.
Inverting the AI Runtime Stack
What if LLMs authored rules instead of making runtime decisions? Nautilus v2 inverts the stack — deterministic execution at runtime, intelligent authoring in the background.
Building a Service Mesh for AI Agents
Agent actions are richer than HTTP requests — tool calls, escalations, handoffs. Bosun governs them at the boundary, before they execute.
Orchestration vs. Choreography in Multi-Agent Systems
When should agents coordinate through a central orchestrator, and when should they self-organize? The answer depends on trust boundaries.