Multi-agent systems where agents with roles hand tasks to each other
When a process is too long for one agent, we split it into roles. One agent takes the task, another gathers data, a third checks the result. An orchestrator keeps the order, and a person can see every step.
What hurts
The process runs through several teams and systems: request, check, calculation, approval, reply to the customer. A single agent with a long prompt gets lost on a chain like this. It forgets early rules and skips steps.
You have already automated parts of it separately, but they do not talk to each other. Tasks get lost at the handovers, and who did what is visible only in email threads. An error at step three surfaces when the customer complains.
A multi-agent system is more complex and more expensive than one agent. It pays off where there are several roles with different rules and access rights. If one agent can do the job, we will tell you so during discovery.
What we build
- Process split into agent roles with separate prompts and permissions
- Orchestrator on Temporal: queues, retries, timeouts, state of every task
- A reviewer agent that checks the result before the next step
- Shared task memory in PostgreSQL without passing extra data around
- Tools for CRM, ERP, email, documents and internal databases
- Human approval points for money, contracts and customer decisions
- Evals for every role and for the whole chain
- A flow dashboard: where each task is and why it stopped
What you get
- Code of the agents, orchestrator and tools in your repository
- A deployed system with monitoring on your servers or ours
- A map of roles, handover rules and responsibility limits
- A test suite for every role and end-to-end scenarios
- Training for the team that approves and supervises the agents
How we measure the result
- Share of tasks that pass the whole chain without human input
- Time for a task to go from the first step to the last
- Number of errors caught by the reviewer agent
- Share of tasks stuck at handovers between roles
Systems and stack
Industries
Questions
Why a multi-agent system instead of one agent?
Each role has a short prompt and only its own permissions. That means less confusion and easier debugging. The reviewer agent checks the work of the others. But the system costs more, so we recommend it only for processes with several roles and different rules.
Why does it start at $15,000?
There are several agents, an orchestrator, and tests for each role and for the whole chain. The price depends on the number of roles, systems and approval points. You get a fixed estimate 2–3 days after discovery. If we exceed it through our fault, we finish at our own cost.
How long does development take?
From 8 to 16 weeks. We launch the system in parts: first one or two roles on part of the flow, then the next ones. There is a demo every week, so you see progress and can change priorities before the next stage.
How do you control agent errors?
Every role has its own tests, and the reviewer agent checks the result before the next step. A person approves decisions about money and contracts. Every step is stored in the log, so any task can be replayed from start to finish.
Where does the system run and who sees the data?
On your own server or on our dedicated server in the EU (Hetzner, Germany). For sensitive data we use our own LLM on vLLM, so data does not go to third-party models. We hold ISO/IEC 27001:2022 and sign an NDA and a GDPR DPA before access. For each role we document what it may do, which also helps with EU AI Act paperwork.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.