Relevance AI Workforce: How to Build a Multi-Agent Canvas

A Relevance AI Workforce is a canvas where several specialist agents, tools and conditions pass one job along, instead of a single agent trying to do everything. You build it from four node types (triggers, agents, tools and conditions) joined by one-way connections, and you decide on each connection whether the agent chooses to hand over or always hands over. Workforces are unlimited on every paid plan, starting with Pro at $19 a month billed annually or $29 a month billed monthly for the whole organization. The canvas itself is not gated. The real limits sit elsewhere: connections only run forward, every tool run spends an Action, and evaluations stay on Enterprise.

Before the first node: plan, browser and the job you are splitting

Every organization now needs a paid plan, and Relevance AI's pricing pages describe no free trial. The builder runs in a desktop browser; the iOS and Android apps are Chat only, so canvas work happens at a desk. Pro includes unlimited agents, tools and workforces with two Build Users and one project, which is enough for one person or a pair to design and run a multi-agent flow.^[1] Team comes with five Build Users, 45 End Users and five shared projects, which matters once other people need to approve tasks or reuse your agents.

Start from the job, not the canvas. A Workforce pays off when one agent would otherwise juggle research, drafting, checking and sending in one long prompt. The docs now point multi-agent builds here and treat the older sub-agents inside an agent as legacy, so do not start new work that way. Build and test each specialist agent on its own first; our guide to Relevance AI agents covers prompts, models, memory and approval modes at the single-agent level.

Lay out the canvas: agents, shared tools and what you can edit there

A new Workforce opens with a manual trigger already placed. You drag agents onto the canvas, and there is no hard limit on how many one Workforce can hold. The same agent can sit in several Workforces at once. That reuse has a catch. Editing an agent from the Workforce builder changes it everywhere, because no Workforce-specific agent settings exist. The builder also hides two things: the agent's own triggers, since the Workforce has its own, and the agent's Advanced Settings. Open the agent itself for those.

Tools are standalone nodes, not buried inside one agent. Several agents can share one tool, and a tool can feed another tool directly, passing its output along without any agent in between. Because tools are independent, you can swap one for another, say Notion for Confluence^[p1], without reconfiguring each agent that used it. When you edit a tool that several agents share, a banner tells you how many agents the change affects; clone the tool if the change should apply to only some of them. Tool-to-tool edges can run automatically or wait for human approval before the second tool processes the data. Our Relevance AI tools guide shows how to build the steps inside each tool node.

Wire the handoffs: AI connection or Next step

Every edge between two agents is one of two kinds. An AI connection lets the first agent decide whether to call the second, guided by a plain-language instruction you write about when to call that agent. A Next step always hands over once the first agent finishes, whatever its state.^[2] Use AI connections where judgment helps, such as routing only qualified leads, and Next step where the order never changes.

Each connection carries three more settings. The approval mode is auto run, approval required or let agent decide; in approval required mode the agent drafts the task and waits for a person. Max auto runs is either no limit or a number, and once that number is reached the agent asks for approval before using the connection again. Set it. It is the cheapest guard against a loop. Task behavior is either create a new task each time or continue the same task.

Continue same task keeps context and history, and the second agent can read and update metadata the first one set; short-term memory is shared only in this mode. Create new task gives each agent its own memory. Parallel handoffs need two things: the sending agent's Parallel Tool Calls setting, still in Beta, turned on, and every outbound connection set to create a new task. Continue-same-task connections run one after another. Trigger-to-agent edges offer the same choices, and one recurring schedule can feed several agents, each with its own settings.

Branch with conditions and decision models

Conditions let a Workforce skip steps or pick a path instead of running everything in the same order. There are two types. LLM-based conditions take natural-language logic and also handle time checks such as whether it is a weekday. Rule-based conditions are deterministic filters with equals, contains, starts with, ends with, exists, empty and the four numeric comparisons. With several filters, Match All needs every one to be true and Match Any needs just one.^[3]

One trap catches most first builds. Rule-based conditions cannot reach nested fields inside an object, so a tool that returns a nested record will not filter cleanly. Have the tool return flat values as separate named outputs instead. A condition wired to a tool works like an IF statement: true runs the tool, false skips it. Several conditions can read the same data at once, so more than one tool may fire in parallel. Plan for that when two branches write to the same record.

Since October 2026, condition nodes and a new Decide tool step can use decision models. These answer multiple-choice, score and yes-or-no questions with a probability for every option, but they do not write text. Jev 1.13 needs your own TypeSafe API key; GPT-6 Luna Decisions runs without your own key in the US and EU regions. Reach for them when an LLM condition's yes or no needs a confidence figure you can route on.

Start it, watch it and pay for it

Beyond the manual trigger, a Workforce can start from integration triggers (email, calendar, CRM, webhook) and recurring schedules: daily, weekly, monthly, custom dates or cron expressions, which should include the year field. Zapier can start one through its Message Workforce event, reaching apps that have no native Relevance AI integration. Schedule triggers cannot be test-activated yet, so test with the manual trigger first. Pause all triggers stops integration, webhook and schedule triggers without deleting them; manual triggers stay live. Pause, cancel and resume controls for Workforce triggers arrived in January 2026.

The Workforce Task View lists active, completed, awaiting-approval and failed tasks. From there you start a manual run, approve, reject or modify agent work, and export conversation logs. An approval nobody answers leaves the task pending until someone approves or rejects it. On paid plans, task alarms send failed, awaiting-approval and escalated tasks to email, Slack or Teams, so route them to someone who checks daily. Every save creates a version you can restore.

Cost follows the tools, not the canvas. Each tool run spends one Action, and a run that fails still counts.^[4] Pro includes 2,500 Actions a month (30,000 a year) plus $20 a month of Vendor Credits for model and tool costs. More Actions come as a top-up pack at $80 per 1,000-Action pack, and unused top-up Actions roll over to the next billing cycle. By our arithmetic, Pro billed monthly plus one pack comes to $109 for a month with 3,500 Actions. A Workforce that chains five tool nodes per task spends Actions five times faster than one agent calling one combined tool, so merge steps inside a tool where order never changes.

Sharing, plan gates and what the canvas cannot do yet

Sharing a Workforce as a cloneable template or turning on any embed option makes it public. Anyone with the link can clone or embed it, and search engines may index it. An embedded Workforce runs on the owner's credits, so treat a public embed as a spending decision. Since May 2026, admins can switch off public cloning and embedding per project. On Enterprise with role-based access, sharing needs permissions on the Workforce and on every agent inside it.

Some limits shape the design. Connections are one-way: the target agent cannot send information back through the same edge, and two-way communication is only planned. In Relevance Chat, a Workforce cannot yet be called together with another Workforce or an agent. Tasks started through the JavaScript SDK accept no file attachments, and Workforces cannot be submitted to the Marketplace. Workforce evaluations arrived in June 2026, but evaluations, Workforce ones included, are an Enterprise feature, so Pro and Team builders test by hand.

The plan matters when a node needs a gated agent type. Calling (phone) agents, meeting agents, the analytics dashboard and bulk scheduling start on Team. The docs suggest putting a phone agent in a Workforce so a parent agent researches before each call, a pattern our Relevance AI for sales guide builds on. Salesforce, Snowflake and Zendesk triggers need Enterprise, even though their tool steps work on every plan.

When a Workforce beats one bigger agent

Choose a Workforce when the job has separable stages with different tools, risks or reviewers. Then per-edge approvals, conditions and shared tool nodes give you control that one long prompt cannot. Stay with a single agent when one specialist and a handful of tools already finish the task, because every extra tool node adds Actions, and every extra node is another place for a run to stall. Begin on Pro with two or three agents, AI connections, a capped max auto runs and approval required on anything customer-facing; move to Team when colleagues need seats, analytics or phone agents.