Friday, 4:47 PM. A small business in Laval gets a call. The customer wants to book an appointment, asks about pricing, then pivots to a billing question about a charge they don't recognize. Three different topics. Three different business logics. And a single AI voice agent on the line trying to handle it all.
This is exactly the scenario pushing more and more businesses to ask: is one agent enough, or do we need several?
The question isn't theoretical. On April 13, 2026, ElevenLabs pushed an update that changes how these systems can be built — and makes multi-agent architecture accessible to SMBs, not just enterprises with a dedicated engineering team. Let's break it down.
Single agent: what it actually means
An AI voice agent in its simplest form is: a system prompt, a model (Claude, GPT, Gemini), and a handful of tools the agent can call. One agent handles the entire conversation, from "hello" to "thanks, have a great day."
It's the default architecture. It's also what works best for 80% of Quebec SMBs — worth saying upfront, because the marketing around "agent swarms" can give the impression you absolutely need to complicate things.
When is single-agent enough?
- Call volume is manageable (under 200 calls/day)
- Topics handled are relatively homogeneous (bookings, info requests, message taking)
- The business scope fits within a 20-50 page knowledge base
- The business doesn't have multiple distinct lines of service
For a hair salon, a dental clinic, or a small accounting firm, single-agent does the job. It's also simpler to maintain, debug, and explain to the team.
Multi-agent: when it starts to make sense
Multi-agent isn't "two agents instead of one." It's an architecture where several specialized agents cooperate within the same call — a triage agent that qualifies the call, then hands off to a booking agent, then to a billing agent if needed. The caller doesn't hear the seam. They just feel like the agent "gets everything."
Technical advantage: each agent has its own prompt, its own tool set, its own focused knowledge base. The billing agent doesn't need to know how the calendar works. The booking agent doesn't need to know refund rules. Result: fewer hallucinations, sharper answers, and business logic that stays readable as the company grows.
The downside: it's more complex to orchestrate. You have to manage context handoff between agents (a caller who already gave their name at triage shouldn't have to repeat it to billing), handle failures, and have a clear view of the "call journey."
What ElevenLabs changed in April 2026
The April 13, 2026 update added three concrete things that make multi-agent practical:
1. The visited_agents array: each conversation response now includes the list of agents that participated in the call. Before, this was a black box. You knew there had been transfers, but reconstructing the path meant digging through logs. Now you have the full trace, agent by agent.
2. Scoped analysis: evaluation criteria can now apply either to the entire conversation, or only to portions where a specific agent was active. This is huge for QA — you can evaluate "did the billing agent explain the fees properly?" without the score being polluted by triage quality.
3. Agent test folders: you can now organize test scenarios into folders, with endpoints for creating, updating, and deleting them. Game-changer when you start having 5 or 6 agents, each with their own test cases.
For technical details, the official ElevenLabs changelog documents the full v2.43.0 release.
A concrete multi-service SMB case
Picture a veterinary clinic in Sherbrooke that does three things: appointment booking, emergencies (with a species-specific protocol), and a specialty food shop. In single-agent, you end up with a 15-page prompt trying to cover everything. The agent eventually mixes contexts, especially under pressure.
In multi-agent, you split it:
- Triage agent: "Hello, is this a booking, an emergency, or a purchase?"
- Booking agent: connected to the calendar via MCP (covered in detail in our guide to live appointment booking with MCP)
- Emergency agent: follows a strict protocol (animal species, symptoms, severity level) and triggers human alerts if red
- Shop agent: has access to inventory and prices
Each agent does one thing, well. And thanks to visited_agents, you can now know, after the fact, exactly what path a call took — useful for post-mortems or identifying where transfers stall.
Vapi Squads, OpenAI Realtime, or ElevenLabs? The quick math
Three platforms dominate this segment in Quebec in 2026, and each has a different approach to multi-agent.
Vapi popularized the concept with its Squads feature. Pro: very clean, designed from day one for multi-agent. Con: pricier and latency is sometimes variable depending on region.
OpenAI Realtime API (with the new gpt-realtime model now GA, and a 20% price drop vs gpt-4o-realtime-preview) is attractive for businesses already running other workflows on OpenAI. At $32 / 1M audio input tokens and $64 / 1M audio output tokens, the math works out fast. Details in OpenAI's official announcement.
ElevenLabs has the edge with the new April 2026 features and — critical for Quebec SMBs — on-premise deployment as of April 9, 2026 (covered in our Law 25 and on-premise analysis).
And what about Law 25?
Question nobody asks enough: when you chain multiple agents, customer data flows between them. The name given at triage ends up in the billing agent. Recent history captured by the booking agent gets shared with the emergency agent. From a Law 25 perspective, this context-sharing is an internal communication of personal information.
It's not a problem in itself — Law 25 doesn't forbid it. But it has to be documented in your PIA (privacy impact assessment) and included in your privacy policy. If each agent has its own provider or infrastructure, it also multiplies the out-of-Quebec transfers to declare. One more reason to consider on-premise if you're going multi-agent.
Practical decision: 4 questions to ask first
- How many distinct "lines of business" does your company have? One = single. Three or more = multi starts to be worth it.
- Do your different flows have different compliance rules? (e.g., medical emergency vs. product sale.) If yes, multi saves you spaghetti prompts.
- Do you have a team to evolve this? Multi-agent needs more governance. If you're managing alone, single-agent stays healthier.
- What's your monthly call volume? Below 1,500 calls/month, multi-agent ROI is rarely justified vs a well-tuned single agent.
For most Quebec SMBs, the honest answer in 2026 stays: start single, migrate to multi-agent when the pain becomes obvious. It's rarely the right starting point for a first implementation, but it's often the right step 12 to 18 months later.
How TECHMA sets this up
When you work with TECHMA to deploy an Agent IA Vocal, we handle the full architecture — single vs multi-agent decision, ElevenLabs configuration (including the new April 2026 features), calendar/CRM integration, Law 25 compliance, and testing. You don't code or configure anything. We almost always start with a well-tuned single agent, and after 60 days of real use, we evaluate together whether multi-agent splitting brings measurable value. No needless complexification.
What to remember
Multi-agent isn't an end in itself. It's a tool that becomes relevant when your business has call flows distinct enough to justify the added complexity. The April 2026 ElevenLabs updates (visited_agents, scoped analysis, test folders) finally make it practical to operate for SMBs — but they don't make it universally necessary.
If you're hesitating, start single. Measure. And let real friction points guide the decision to chain later. Often, a well-briefed single agent with a properly structured knowledge base, like we covered in our knowledge base search analysis, performs far better than a poorly orchestrated multi-agent setup.