In this article
Single agent or multi-agent: what to ask your development partner
Choose the simplest architecture that completes your tasks with appropriate controls. Multiple agents need a measurable advantage, such as handling independent work, rather than a separate agent for every business department.

Start with the task and the evidence you need
To choose between a workflow, a single agent and multiple agents, request a comparison on your own cases. The selected architecture should meet an agreed quality threshold, respect permissions and remain manageable for your team.
A deterministic workflow follows defined steps and conditions. A single agent chooses tools and adapts its steps within one working context. A multi-agent system distributes some work across separate loops with their own instructions or contexts. A chatbot can provide the interface for any of these approaches.
For a buyer, every additional layer should address an identifiable problem.
October’s signal: simplification deserves a test
In its October 22, 2026 article about folding subagents into a main loop, Anthropic describes deployments where one loop with tools and skills improved outcomes. It also retains uses for subagents, including parallel work and context limitations. This vendor experience does not establish that a single agent is universally superior.
Our procurement recommendation is to ask your partner to compare a simpler alternative with its proposed design. An impressive diagram can become difficult to debug when agents exchange incomplete summaries.
Compare three approaches on one process
Consider a customer service request: identify an order, check its status and prepare a response. This is an illustrative example.
| Observed situation | Reasonable starting point | Reason to add complexity |
|---|---|---|
| Structured data and stable rules | Workflow with explicit conditions | Exceptions need useful interpretation |
| Ambiguous requests and a few tools | Single agent with controlled actions | Evaluations expose a limitation |
| Genuinely independent research | Parallel subtasks with synthesis | Benefits exceed communication and operating costs |
| Conflicting or sensitive sources | Human review and access rules | Technical separation actually protects the data |
A different department does not automatically require another agent. If logistics and billing need the same case details, splitting them may increase information loss. Conversely, analyzing independent documents can make separation useful.
Separating context is not automatically the same as separating permissions. The relevant tools and services still need to enforce access controls.
Attach an evaluation plan to the proposal
Build an initial case set with the business team, using anonymized examples or authorized test data. Include common requests, missing information, a contradiction, an out-of-scope request and a prohibited action. Keep a separate final verification set so improvements are not limited to familiar examples.
For each case, define an acceptable outcome, the expected source and the permitted action. Some correct outcomes are escalations: asking a question or handing over to a person can be part of successful delivery.
Compare architectures under equivalent conditions: documented model versions, the same tools and permissions, and the same maximum spending budget. Repeat runs to observe variation. An early sample reveals defects; it cannot certify production reliability.
Track at least:
- correctly completed cases, using an agreed definition;
- unauthorized actions and invented information;
- completion time, including slower cases;
- cost per accepted case, including retries;
- human correction effort and ease of investigation.
Faster execution that increases correction work may leave the team worse off.
Plan for failures before dividing the work
Suppose the billing tool times out. The system needs a rule for continuing, waiting or escalating. A retry should also avoid creating a duplicate write.
Request time limits, spending limits and a named exception owner. Logs should make a case traceable while avoiding unnecessary exposure of its data. Permissions should match useful actions, with approval before sensitive operations.
These requirements apply to every architecture. They prevent a single agent from becoming an uncontrolled system and stop a multi-agent design from being mistaken for sufficient supervision.
What a development partner should deliver
The proposal should explain each component, provide comparative results and identify what your team will maintain. Ask how to change a model, update a business rule and return to manual handling.
MTECH builds AI agents connected to business systems and workflow automation from Paris. We recommend funding evidence on the process first, then choosing an architecture that supports it.
To review a proposed design, share your workflow and acceptance criteria. The useful discussion concerns the cases the system can complete correctly and the work your team can recover when it reaches a limit.
FAQ
Frequently asked questions
Is a multi-agent system always more effective?
No. Multiple agents can help with independent tasks or distinct context requirements. They also introduce communication, latency and failure points. Compare alternatives on the same cases before choosing.
Is a chatbot different from a single agent?
A chatbot mainly describes a conversational interface. It can connect to one agent, multiple agents or a fixed workflow. The visible conversation does not establish the underlying architecture.
What should a multi-agent development proposal explain?
Ask why each agent is necessary, which data it receives, which actions it can perform and what measured benefit justifies its cost. Include end-to-end evaluations and a recovery procedure.