Giving agents a map: the teamwork graph

Ask an agent to “get sign-off on the Q3 launch” and it will happily generate a plausible plan. Whether that plan touches the right people depends on something the model doesn’t have: a model of who does what, and how work connects. That missing piece is a graph.

Context windows aren’t org charts

The instinct is to stuff more context into the prompt — more docs, more history, more people. But an organization isn’t a pile of documents, it’s a network: people own systems, systems depend on other systems, decisions need specific approvers. A flat context window flattens exactly the structure the agent needs.

Multi-hop is the whole game

The interesting questions are almost never one hop away:

“Who needs to approve a change to the billing service?”

Answering that means traversing: billing service → owning team → team lead → their current delegate while they’re on leave. A graph makes that a walk. A prompt makes it a guess.

Exposing the graph everywhere

A graph is only useful where the agent already works. That’s why it’s worth exposing the same reasoning layer across surfaces — a web UI for humans, an MCP server for other agents, a CLI for scripts, a browser extension for in-context lookups. The graph is the product; the surfaces are just doors into it.

This is a starter post — replace it with your own writing.