Agent systems practice

Multi-agent systems: specialists under one supervisor

When one agent can’t hold the whole job, we build specialists under a supervisor, each with a contract. Week 1 ends with a plan and a price, in writing.

A round paper table with a tall vermilion-tipped column at its centre, joined by vermilion strips to four different paper shapes, a marigold bell at its rim.

What the service covers

We start with the contract: which sub-tasks are real, which roles they need, and what each role refuses to do.

  • The topology and its contracts

    Who does what, and a contract per agent: its instructions, its tools and when it escalates.

  • A supervisor that owns the run

    One agent holds the conversation and shared memory, and calls the specialists.

  • Specialists with narrow jobs

    Each does one thing, with its own tests and reference tasks drawn from your work.

  • An end-to-end evaluation

    We score the whole system, so a run where every agent passed but the result failed gets caught.

3 ways to engage

Pick the one closest to where you are. Week 1 settles the scope, in writing.

  1. 1What you get

    Multi-agent design review

    A written answer on whether you need more than one agent, before you commit. Often the answer is no.

    How it runs

    One week of discovery that ends with a plan and a price, in writing.

  2. 2What you get

    The multi-agent build

    Design, build and evaluation, handed over with the topology document and a register of failure modes, so nobody redesigns from scratch.

    How it runs

    Fixed scope and fixed cost, agreed before work begins, with a named person on every gate.

  3. 3What you get

    Run it with us

    We keep operating the system with your team and add a specialist only when a measured bottleneck calls for one.

    How it runs

    In your accounts, behind the same gates. A documented handover whenever you want to take it over.

Who the service is for

Multi-agent systems / Fit check

It fits

  • Teams running long work, such as legal review, claims or research, where one agent has proved not to be enough.
  • Engineering teams whose agent in production now runs out of context or expertise.
  • Operations leads automating a process that already spans several roles.

It doesn’t fit

  • Choosing multi-agent because it sounds advanced. Start with one agent and a list of where it falls short.
  • Demos. Without a real bottleneck, more agents only multiply cost and failure.
  • A first agent. Our agent systems practice starts with one, measured before anything grows.

Questions about this service

Every engagement runs the same method: an agent proposes, a person approves.

When is multi-agent the wrong answer?
When one agent and a checklist would do, when the tasks aren’t really different, or when nobody can say why one agent falls short. In those cases we recommend a single agent and say so in writing.
Does it cost more than a single agent?
To run, yes: every agent adds model calls to each request. In return the system can finish work one agent can’t. The build itself is fixed scope and fixed cost, with the price in writing in week 1.
How do the agents share context?
A shared store the supervisor reads and writes, often Postgres or Redis; messages with explicit hand-offs; or a mix, where each specialist keeps its own state. The shape of your work decides which.
What happens when agents disagree?
It is written into the contract. The supervisor decides by rules, a critic agent scores both answers, or the question goes to a person. We design the failure path first: a clean escalation beats a silent disagreement.
When does a person take over?
The triggers are set before the build: low confidence, conflicting answers, a missing tool, a timeout. Each one sends the work to a named person. Escalation is part of the architecture, never a patch added later.