Your company already runs the hardest parts of an AI agent platform

Short summary

Companies already have the infrastructure for AI agent platforms: org charts for scoping, directories for identity, and communication channels for human feedback. The author proposes a three-tier context architecture — a root constitution, project constitutions, and on-demand documents — where cost runs opposite to volume. Two human channels (escalation and calibration) should run through existing tools like Teams, with corrections folded back into project-level rules.

• •Use org-chart inheritance for agent context: root constitution, project constitution, on-demand docs — smallest tier is most expensive
• •Escalation (agent asks human) and calibration (human corrects agent) are two distinct channels both needed for production agents
• •Route human channels through existing tools like Teams; beautiful dashboards go unread

Generated with AI, which can make mistakes.