Agentic / AI
Ganglion:Agentic AI Frameworks
A multi-agent system built to find out where AI actually belongs.
AI proposes. A rule or a person disposes. As an answer repeats and holds up, it hardens into a rule or a line of code, and the model's share of the work shrinks from there. That's the direction this is built to move in: less AI over time, not more, applied exactly where it earns its place.
It starts read-only. Point a model at a process before ever letting it act on one, and test write access as its own separate step. At work, that meant piloting a read-only Claude connection to Business Central first, testing write access for the next stage only after, and handing IT a specification before anything changed.
The same discipline produced something small and concrete: a monthly tool, built with Claude Code, that flags the known causes of fluctuation in declared revenue, used directly by the Salesforce team. It's deterministic everywhere it can be, and only asks a model to look at what's genuinely ambiguous.
This thinking also helped shape an internal AI knowledge-sharing group before it was formalized, and continued as something it consulted afterward.
Ganglion is a multi-agent software team, built on LangGraph, to test directly where AI actually belongs in a workflow instead of guessing from the outside.
The agents don't talk to each other, and they don't share a memory. What connects them is the work itself: a shared substrate of artifacts sitting between roles, with a controlled channel for passing a finding forward, not a live conversation. Coordination comes from local rules, the way a flock holds its shape without a leader, not from one agent directing the others.
The interesting part was never any single agent. It's the pattern repeating: the same simple unit, doing the same bounded job, in enough places that the structure starts to look less like a tool chain and more like a small organization.
It's the same principle as the rest of this page, run as an experiment: let a model propose inside a narrow, isolated role, and keep the actual deciding deterministic, outside of it.