Skip to content
dhii.ai

Agentic AI & Automation

Multi-Agent Systems

Several specialised agents — a planner, researchers, a reviewer — coordinated by explicit orchestration, for work too broad for a single prompt.

Architecture & baseline
2–3 weeks
Production system
6–12 weeks
Typical team
2–3 engineers

Honest framing first

Multi-agent systems are fashionable, and fashion produces over-engineering. A team of five chatting agents makes a compelling demo, but in production each extra agent adds latency, cost and a new place for things to go wrong. Our default position is that you should earn the right to add an agent by showing, on a test set, that it beats the simpler design.

That said, some problems really do benefit. When a job spans many sources, demands different expertise at different stages, or is simply too large for a single context window, splitting it among specialists — with a clear coordinator — produces better results and is easier to maintain.

Patterns we use

  • Planner–executor. One agent drafts a plan of steps; executor agents carry them out with narrow tools; the planner revises when results come back. Good for research and data-gathering tasks.
  • Supervisor with specialists. A routing agent decides which specialist — billing, technical, legal — should handle each part of a request. Good for broad internal assistants.
  • Generator–critic. One agent produces, another reviews against explicit criteria and sends it back with specific objections. Good for proposals, reports and code.
  • Pipeline with checkpoints. Fixed stages with an agent at each, and validation between stages. The most predictable pattern and often the right one.

We implement these as explicit graphs in code, not as open-ended conversations. Each node has typed inputs and outputs, the shared state is a defined schema, and every transition is logged.

What we watch in production

Multi-agent failures tend to be quiet: a hand-off loses a key detail, a reviewer approves something it shouldn’t, a loop burns tokens for minutes. We instrument each agent separately so you can see quality per step, and we set budgets so that a run that exceeds its allowance stops and escalates instead of carrying on. Over time the traces show which agents add value and which can be merged or removed — and we act on that.

Deliverables

What you get

  • 01

    Architecture decision record

    A written case for why the task needs multiple agents — or a simpler design if it doesn't.

  • 02

    Orchestration graph

    An explicit state machine of agents, hand-offs, shared state and stop conditions, versioned in code.

  • 03

    Specialist agents

    Each agent with a narrow role, its own tools and its own evaluation set.

  • 04

    Supervisor and critic loop

    A reviewing step that checks outputs against requirements before anything leaves the system.

  • 05

    Budget and loop guards

    Hard limits on steps, tokens, time and spend per run so no job runs away.

  • 06

    Trace viewer

    Per-run visualisation of which agent did what, with costs, so failures can be pinned to a step.

Process

How a multi-agent systems project runs

  1. Step 1

    Decompose

    Break the target task into sub-tasks and check which genuinely need separate reasoning contexts.

  2. Step 2

    Baseline

    Build the simplest single-agent version first and measure it — this is the bar to beat.

  3. Step 3

    Orchestrate

    Introduce specialist agents only where they measurably improve quality or cost.

  4. Step 4

    Stress-test

    Run adversarial and long-horizon cases to find loops, hand-off failures and cost spikes.

  5. Step 5

    Deploy

    Ship with tracing, budgets and alerting, then tune on real traffic.

Engagement models

Ways to work together

Every engagement starts with a fixed, written quote. See engagement models & pricing or the how much does it cost to build an ai agent?.

FAQ

Multi-Agent Systems: common questions

When do I actually need a multi-agent system?

When a task has clearly separable parts that each need different context, tools or expertise — for example research across many sources, then synthesis, then compliance review — and when a single agent's context gets too crowded to perform well. For many business workflows a single well-designed agent is cheaper and more reliable.

Which framework do you use?

We most often use LangGraph or plain Python state machines because explicit graphs are easier to test and debug. We also work with CrewAI, AutoGen and the OpenAI Agents SDK when a client is already invested in them.

Aren't multi-agent systems expensive to run?

They can be. Every agent call costs tokens and time. We cap steps and spend per run, cache shared context, and route simple sub-tasks to small cheap models. We report cost per completed task so you can judge value.

How do you stop agents looping or arguing forever?

With explicit stop conditions, step limits, a supervisor that can terminate a run, and deterministic hand-off rules. Free-form agent conversation is a demo pattern, not a production one.

Related services

Start a project

Have a workflow that should run itself?

Tell us what you want automated or built. You get a written assessment of feasibility, risks and a fixed quote — usually within two business days.