A Multi-Agent System: 10 Agents Coordinating Work Previously Done by Hand
An IT services firm whose internal delivery workflow had outgrown what people could carry by hand. The work ran across several internal teams, with each team handling a part of the same process without visibility across the entire workflow.
An IT services firm had an internal delivery workflow that had outgrown what teams could manage manually. Instead of building a single automation pipeline, Orbis developed a multi-agent system with 10 agents coordinating work across teams. The system now automates most of the workflow and brings people into the process only where human judgement is required.
Ten agents coordinating work across teams
The workflow spanned multiple teams, each handling part of the same process. The system now automates most of the workflow and brings people into the process only where human judgement is required.
A recurring workflow touched multiple teams
A recurring workflow touched multiple teams, with each team performing a narrow part of the process manually. Information was pulled from different sources, reformatted, and passed from one team to another.
As the business grew, the workflow grew with it, while the coordination cost increased faster than the work itself. The challenge was not simply to automate one process, but to build a multi-agent system capable of coordinating 10 agents across multiple teams.
The workflow split into agent teams, not a single flow
Scoping revealed that the workflow naturally split into agent teams rather than a single flow. Each team could own a specific responsibility, but the boundaries between teams and the shared state they depended on created the main architectural complexity.
As a result, the engagement combined AI Build with AI Consultancy, with architecture decisions being as important as the implementation itself.
Agent teams, orchestration and human-in-the-loop gates
One system replaces the hand-offs between teams
One system now replaces the hand-offs that previously existed between teams. The workflow flows through the agent teams and surfaces to a person only where judgement genuinely does not transfer. Hours recovered, turnaround, and error rate are all reported against the same baseline used to evaluate the system.
The hardest part was the shared state. Each team had its own understanding of what a unit of work looked like, and the orchestration layer had to reconcile these differences before the agents could reliably use the shared state. This complexity was not obvious in a single-agent demonstration but became decisive in the multi-agent system.
“The client described the workflow as something the teams had been carrying for years and had come to treat as part of the normal cost of doing business. With the new system, 10 agents now run the workflow, while people step in only where judgement matters. The organization can measure the hours recovered each week, and the documentation was detailed enough for the internal team to understand how the system was built.”
