# Relevance AI: How to Build a Multi Agent Workforce

*How-to — 2026-10-24 — by Mahmoud Zalt*

How to build a multi agent workforce in Relevance AI: putting agents on the canvas, configuring handoffs, conditions, approvals and escalations.

**Short answer.** **A Workforce in Relevance AI is a visual canvas where you drag existing agents in and draw connections between them.** Each connection is configured as either an AI connection, where the first agent decides when to involve the second, or a forced next step that always hands over. You also choose whether the receiving agent starts a fresh task or continues the same one, which is the setting that decides whether context survives the handoff.

**About this guide.** This guide is based on publicly available information and Relevance AI's own documentation, checked on 17 August 2026. Sistava is not affiliated with Relevance AI. Interfaces and pricing change often, so confirm the current steps in their documentation before you rely on them.

The moment people reach for a multi agent setup is usually the moment a single agent has started doing four jobs badly. Splitting the work is the right instinct. Where it goes wrong is the handoffs, and Relevance AI puts almost all of its configuration exactly there.

Before you open the canvas, write the process down as a sentence per stage. If you cannot describe the handoff in words, you will not be able to configure it, and the canvas will happily let you draw a line that means nothing.

## When is a workforce actually better than one agent?

When the job has genuinely separate stages that need different instructions or different tools. Their own guidance on building agent teams is to think in tasks rather than roles, and to give each agent a narrow goal, because an agent with a broad remit and limited context produces inconsistent choices.

Their workforce documentation makes the same point from the other direction: simpler workforces with clearly defined roles tend to perform better than complicated ones. Two agents with a clean boundary beat five with overlapping responsibilities, every time.

The example their team writing gives is a content pipeline: a manager that takes the request and routes it, a researcher with search and scraping tools, a writer, and a specialist that handles formatting. Each has one job, and the manager holds the coordination.

## How do you create the workforce?

You build the agents first, then assemble them. The canvas does not create agents for you, so anything you want on it has to exist in your workspace already, tested on its own in chat.

### Assemble a workforce on the canvas

1. **Build and test each agent separately** — Every agent should already do its own job correctly in a normal conversation. A workforce does not fix a weak agent, it just gives that agent colleagues who now depend on it.
2. **Open Workforce and start a new one** — Workforce sits in the left sidebar. Choosing New Workforce opens the visual canvas you will build on.
3. **Drag your agents onto the canvas** — Agents come from the sidebar. Place them in the order the work actually flows, because the layout is documentation as much as configuration and other people will read it.
4. **Draw the connections** — Connect agents by dragging from the output node of one to the input node of the next. A connection line appears and the edge settings panel opens, which is where the real decisions get made.
5. **Configure each connection** — For every line, choose an AI connection or a forced next step, and decide whether the receiving agent starts a new task or continues the existing one. Do not leave defaults you have not read.
6. **Add tools and conditions where the flow branches** — Tools can sit on the canvas as their own nodes, and conditions act as decision points that route work down different paths. Use a condition rather than asking an agent to remember a rule.
7. **Add a trigger to start the flow** — Triggers are the starting points that begin a workflow. Add one only after the path behaves correctly when you start it by hand.
8. **Save and run it end to end** — Run one realistic request all the way through and read every stage. Most workforce faults are not agent faults, they are handoff faults.

Notice how little of that is about the agents themselves. On the canvas the interesting object is the edge, not the node. Two competent agents joined by a badly configured connection produce worse output than one mediocre agent working alone.

## What are the two ways one agent can hand off to another?

Their agent to agent documentation describes two communication styles, and choosing between them is the most consequential decision on the canvas.

- **AI connection**: the first agent decides for itself when to involve the second. You have to fill in a field describing when to call that agent, and the quality of that sentence is what determines whether the handoff happens at the right moment.
- **Next step, also described as a forced handover**: the first agent always passes control to the second when it finishes. There is no judgement involved, which is exactly why it is the safer default for a fixed pipeline.

Use a forced next step wherever the process genuinely always runs in that order. Save the AI connection for real branches, where sometimes you need the specialist and sometimes you do not. Every AI connection you add is another judgement call the system can get wrong.

## New task or same task, and why does it matter?

This is the setting that decides whether context survives the handoff, and it is the one people get wrong first. Their documentation describes creating a new task, where the second agent starts fresh with its own context, or continuing the same task, where it keeps the context, conversation history, and short term memory.

Continue the same task when the second agent needs to know what was already said, such as a writer following a researcher. Start a new task when the second agent should judge the work independently, such as a reviewer that would be biased by seeing the reasoning that produced the draft. A clean context is sometimes the point.

**Handoffs are one way.** Their documentation states that only one way communication is supported and that back and forth conversation between agents is not yet available. So design your flow as a pipeline, not a discussion. If agent B needs something from agent A, that information has to travel with the handoff, because B cannot turn round and ask.

That constraint is worth designing around rather than fighting. Put the agent that gathers information before the agent that needs it, and make the first one produce a complete enough output that the second never has to ask a follow up question.

If maintaining that diagram sounds like a job in itself, that is a fair reading, and it is the part we approached differently at Sistava. You hire AI Employees for roles and they coordinate as a team without you drawing the edges, so the thing you maintain is the brief rather than the wiring. It is a different philosophy, not a better canvas.

## How do you keep a human in the loop?

Through approvals and escalations, and they are configured in the same edge settings you already have open when connecting an agent to a tool or to another agent. Their documentation describes three modes for how much freedom the connection carries.

- **Auto Run**: the action executes without asking. Their documentation frames this as appropriate for low risk routine work, and notes you can set maximum auto run limits as a guardrail.
- **Approval Required**: the agent drafts what it intends to do and waits for a human decision. Appropriate for high stakes actions and anything that commits resources.
- **Let Agent Decide**: the agent proceeds when confident and asks when uncertain. This one needs written instructions describing the decision logic, otherwise you are guessing what confident means to a model.

The Workforce Task View is where these land. Their documentation describes it as the central place to review pending approval requests, inspect the action an agent is proposing, and read the reasoning behind it. Escalations surface there too, and can be routed onward through tools such as Slack or email to reach a specific person.

One caveat matters operationally. Their documentation is explicit that an unanswered approval stays pending until it is approved or rejected, and that tasks do not proceed on their own. That is the correct behaviour, and it also means an approval queue nobody watches is a workforce that quietly stopped working. Decide who checks it before you turn the trigger on.

## What goes wrong when you scale past two agents?

- **Too many agents, too early**: their own guidance favours simpler workforces with clear roles. Start with two, get the handoff right, and add the third only when a real gap appears.
- **Overlapping responsibilities**: if two agents could both plausibly handle a request, the routing becomes a coin flip. Boundaries between agents should be obvious in one sentence.
- **Everything on AI connections**: each one is another decision the system can get wrong. Forced next steps for the fixed parts of the flow remove whole categories of failure.
- **Context dropped at the handoff**: choosing a new task when the receiving agent needed the history produces confident, uninformed output. This is the single most common cause of nonsense at stage two.
- **Approvals nobody owns**: pending requests do not time out into action, so an unattended queue stalls the flow silently rather than loudly.
- **Cost multiplied by design**: an Action is counted whenever an agent runs a tool, so a four agent chain that each call tools costs more per request than one agent doing the same work. Usage alerts and limits are worth setting first.

Relevance AI gives you a genuinely good picture of your process, and the canvas is a real advantage when you need to show a colleague how the work flows. The cost is that the process is now a thing you own and maintain. At Sistava we bet that most teams want the outcome rather than the diagram, so you hire the roles and they sort the handoffs out between themselves. Both bets are defensible, and yours depends on how much you want to see.

## FAQ

### Do agents in a Relevance AI workforce talk back and forth?

Not currently. Their agent to agent documentation states that only one way communication is supported and that bidirectional back and forth conversation is not yet available. Design the flow as a pipeline where each handoff carries everything the next agent needs.

### What is the difference between an AI connection and a next step?

An AI connection lets the first agent decide when to involve the second, and requires you to describe when that agent should be called. A next step, also described as a forced handover, always passes control on completion. Use forced handovers for fixed pipelines and AI connections only for real branches.

### Does the second agent see what the first one did?

Only if you configure the connection to continue the same task, which preserves context, conversation history, and short term memory. Choosing to create a new task gives the second agent a clean start, which is useful for independent review and harmful when it needed the history.

### How many agents should a workforce have?

Fewer than you think. Their documentation notes that simpler workforces with clearly defined roles tend to perform better than complex ones, and their guidance on agent teams is to give each agent a narrow task rather than a broad human style role. Start with two and grow only where a boundary is obvious.

### Where do I approve an action an agent wants to take?

In the Workforce Task View, which their documentation describes as the central dashboard for reviewing pending approval requests along with the proposed action and the agent's reasoning. Escalations appear there as well and can be routed onward through tools such as Slack or email.

### What happens if nobody responds to an approval request?

It stays pending. Their documentation is clear that tasks do not proceed without explicit authorisation, so the work simply waits. Assign an owner for the approval queue before you attach a trigger, otherwise the workforce stops without anyone noticing.

## Sources

- [Relevance AI documentation, Create a workforce](https://relevanceai.com/docs/workforce/create-a-workforce)
- [Relevance AI documentation, Agent to agent configuration](https://relevanceai.com/docs/workforce/workforce-features/agent-to-agent-configuration)
- [Relevance AI documentation, Approvals and escalations](https://relevanceai.com/docs/workforce/workforce-features/approvals-and-escalations)
- [Relevance AI documentation index of all pages](https://relevanceai.com/docs/llms.txt)
- [Relevance AI blog, Building an AI agent team](https://relevanceai.com/blog/building-an-ai-agent-team)
- [Relevance AI documentation, Pricing](https://relevanceai.com/docs/get-started/pricing.md)

**Tags:** relevance-ai, multi-agent-systems, ai-workforce, agent-handoff, no-code-ai