An agentic room-booking platform and the agents that build it

Members of a coworking network book meeting rooms on a platform we built on a dedicated server, with our own agentic builder. The heart of the system is the agentic layer on top, which lets the network's teams change the platform themselves, without calling a vendor.

  • For members: one API behind every screen, so an AI agent books under the same rules as the button; bookings and cancellations are recorded in the network's back-office, and signed-in members have their own assistant.
  • For the network's teams: each change is prepared by an operations agent briefed in plain language, checked by an independent reviewer against the business rules and tests, approved by the team, and kept reversible.
  • Reliability: prepaid balances that can never go negative, business rules encoded as automated tests, scheduled checks and backups, guardrails before any agent action.
  • The builder: an orchestrating agent splits the work into tracks, each agent with a brief, a scope and a finish line, reading the project state first and recalling memory on a precise question; a read-only adversarial reviewer challenges the build until nothing blocks; production waits for a named human.

Learn these patterns in our Claude & Claude Code training.

Tech Stack

Claude Code, Python / FastAPI, Next.js, PostgreSQL (double-entry credit ledger), workspace back-office API, agent connector (MCP), Dagster, Git release pipeline
A person and an agent feed one booking list, one room slot is confirmed in a grid, and a central agent hub connects to a shield check and a validated person

The challenge

A coworking network opens its meeting rooms to outside companies, who pay with prepaid credit wallets. Every action on screen also had to be available to an AI agent, every rule adjustable without development, and the platform run by the network's own teams after handover, without calling a vendor for each change.

Result

One set of rules for people and AI agents, a platform the network's teams evolve without a vendor, and a build reviewed by an independent agent, with production left to a human decision.