Multi-Agent Development · Specialised agents, one system
theBlue.ai builds multi-agent systems for companies: several specialised agents, each responsible for one part of a process, coordinated by an orchestrator that routes the work, passes on context and hands decisions to people where they belong. The architecture follows your process and runs inside the systems you already have, on-premise, in the cloud or hybrid.
Proof
theBlue.ai has delivered more than 100 AI and automation projects since 2016. Two of our systems split a process across specialised agents behind one orchestrator, and both are described in full in the case studies below.
Some of the organisations we have built for
What the agents take over
Several agents pay off when the work splits into distinct roles or knowledge areas. Each card shows one of those situations and where it comes from.
A consultant uploads a policy paper or a legislative proposal and gets back a structured summary, the key takeaways and answers to follow-up questions. Document analysis and research run as separate agents.
−70% routine workRPP Group
A drafting agent writes strategic plans, core messages and first drafts in the tone the team uses with stakeholders. Quality agents check the output against internal standards before it reaches a client.
5 agents, 1 orchestratorRPP Group
Citizens ask in everyday language. A main agent works out which of eleven domains the question touches and delegates to their agents, calculates permit deadlines and asks back when a request is ambiguous.
11 domain agentsPublic authority, under NDA
A request is classified, routed, drafted and checked before anyone answers it, across email, the ticket system and two more systems. The agents carry the case from the first document to the finished result and hand back what needs a human decision.
Human decisions stay humanTypical starting point
How it works
Orchestratorreads the request, splits the work
Specialist agentsone role each
Quality checkagents that check agents
Your teamdecides and sends
Illustrative example. Figures from the RPP Group case study.
Architecture
The model matters less than the architecture around it: how tasks are routed, how context is passed on, and who is responsible for each step. The process decides which pattern fits, and the number of agents follows from that.
A main agent routes every task to the specialists. High control, with the orchestrator as the point everything depends on.
Our projects so farAgents hand work to each other directly. More resilient, and harder to observe and debug for your IT.
Coordination where control matters, direct hand-offs where they are safe. Increasingly common in enterprise practice.
Two further layers decide whether a system holds up in production: memory, meaning a working memory for the running process plus persistent knowledge about your company, and standard interfaces such as the Model Context Protocol (MCP) that connect agents to your data and tools. More in our article on architecture, integration and governance.
Control
An enterprise agent system does not need to be as autonomous as possible. It needs to work reliably, transparently and under control.
Every agent has a defined task, the data it may use and the tools it may call. The orchestrator knows who is responsible for each step.
Quality agents verify output against your internal standards, so an inaccuracy is caught before it reaches a client.
The system hands back what needs a human decision. Where those points sit is fixed in the process analysis.
You can trace which agent acted and on which data. For RPP Group everything runs GDPR-compliant inside their own infrastructure.
Case studies
Two systems that split a process across specialised agents, and the policy platform one of them was built with.
Public affairs · RPP Group
Senior consultants spent hours on document analysis, data structuring and drafting, work that crowded out the strategic side. ChatRPP now carries it, in the firm’s own tone and inside their own infrastructure.
Read the case study
Government · Polish institution, under NDA
The old chatbot asked citizens to pick a category before asking, and most did not know which one fitted their question. Specialised agents with retrieval now route each question themselves and match plain language to the technical documentation.
Read the case study
Public affairs · Policy-Insider.AI
Public affairs analysts tracked millions of policy documents across countries and languages by hand. We built the AI layer that monitors, analyses and alerts, from first prototype to market release.
Read the case studyHow to start
We start with the process as it actually runs: its branches, exceptions and human decision points. Only then is it clear where an agent is needed, where conventional software is enough and where a person keeps responsibility.
Tell us which process crosses the most roles, desks or systems. We come back within one business day with an initial assessment and a proposal for a 30-minute scoping call.
Describe your processWorkflows mapped with their exceptions, the systems and permissions involved checked, and an architecture proposed with the roles of each agent, scope, timeline and cost. A standalone engagement with no commitment to proceed.
See how we workFAQ
A multi-agent system splits a process across several specialised AI agents, each responsible for one type of task, such as document analysis, research, drafting or quality checks. An orchestrator routes the work between them, passes on context and hands decisions to people. theBlue.ai builds these systems inside existing enterprise IT, on-premise, in the cloud or hybrid.
When the work splits into distinct roles or knowledge areas. The citizen chatbot we designed covers eleven domains with one agent each, because no single agent could hold all of them reliably. When one clear task runs through a few systems, a single agent is the better choice; see AI Agent Development. The number of agents is not the goal. It follows from the process.
Through an orchestrator. It reads the request, decides which specialists are involved and routes the tasks to them. Context is passed on between steps, intermediate results are validated, and exceptions are escalated to a person. In ChatRPP a main orchestration agent routes each query to document analysis, content drafting, strategic planning or policy research.
The one the process calls for. Our systems so far use one orchestrator in front of specialised agents, which gives high control. Peer-to-peer structures are more resilient but harder to observe, and hybrid models combine both and are increasingly common in enterprise practice. The choice is made in the process analysis.
Every agent gets one defined role, with the data it may use and the tools it may call. Quality agents check the output of the others, the points where a person decides are fixed in advance, and it stays traceable which agent acted on which data. Governance is part of the architecture from the first line of code.
The systems you already run, whether standard software or built in-house: ERP and CRM, document stores, internal knowledge and external data sources, each through its interface. ChatRPP answers from internal documents and external policy data. In the citizen chatbot, domain experts maintain the knowledge base themselves in editable lists, without a developer.
Yes. ChatRPP processes confidential political documents GDPR-compliant inside RPP Group’s own infrastructure, so the material never leaves their controlled environment. theBlue.ai deploys on-premise, in the cloud or hybrid, depending on your security requirements.
Rarely in the model logic, and almost always in fragmented data landscapes. A multi-agent system is only as good as the data it can reach, so integration with the systems that hold that data is the real engineering work. That is why the process analysis maps data flows and interfaces before anything is built.
The process analysis has a fixed price from €3k and ends with an architecture proposal that states the roles of each agent, scope, timeline and cost. The build is priced in milestones, and first working components typically arrive six to eight weeks in.
Describe the process and we’ll come back within one business day with an initial assessment and a proposal for a 30-minute scoping call.
