Your agents are only as deep as the graph you give them

The new motion is easy to demo. An agent in Slack, a skill in an IDE, a bot that “does outbound.” The demo searches the web, invents a plausible reason, and writes a paragraph that sounds like every other paragraph. The team claps, and then a rep sends it, and then nobody answers, and the postmortem blames the model.

The model was not the thin part. The graph was. An agent that can call Priorread can ask who is in this room, why this company matched, which tie is real, and whether the opener’s source is still fresh. It can also be told no. Permissions follow the person who connected the key. Lookups are metered. The audit line is dull and that is the point.

We ship an MCP server, a small set of skills, and an API that does not try to be clever. search, profile, connections, opener, watch. If your agent needs a seventh verb, the API is there. If your security team needs the agent to see less than a VP sees, that is a setting, not a favor.

The AI age did not make directories deep. It made shallow data faster. Speed, applied to a title and a territory, is just a quicker way to be ignored. Give the agent the room.