The agent orchestration layer beneath a chat-run business operating system

The startup's product lets a small business run operations through one chat: messaging, documents, time tracking, telephony and more, each a separate system with its own API. A request such as "chase the three invoices that are overdue and book a call with the one that replies" touches several of them in sequence and needs a decision at each step.

A single model call with a long prompt could not do this reliably, could not explain what it had done, and could not be stopped before it did something irreversible. The product needed a layer that mapped intent to capability, carried context across sessions, and knew when to ask a person.

We built the orchestration layer beneath the conversation. An agent registry maps each request to a capability. Skill-based sub-agents carry the work through MCP tool servers, one per connected system and each scoped to least privilege. A custom router sequences the sub-agents, decides what runs in parallel, and recognises when a step needs a human. Persistent memory carries context across sessions and channels.

A persona layer lets a team define an assistant once, with its role, tone, tools and permissions, and publish it company-wide. Every action is logged and replayable. Anything irreversible, out of scope or ambiguous waits for approval in the same chat. Adding a capability is a registry entry and adding a system is a tool server, so the product grows without rebuilds.

Plain-language requests now become finished work across the connected systems, with an approval in the chat wherever a step is irreversible. The startup's own engineers add capabilities as registry entries and systems as tool servers. A proof of concept over two or three systems landed in a few weeks; the production layer with registry, memory, personas and approvals followed inside the usual eight to fourteen week window.