ElevenLabs Agent Workflows is a visual conversation logic builder within ElevenAgents. It allows teams to define the paths a voice agent can take through a conversation — what the agent does when a caller asks about billing, what happens when a caller is verified versus unverified, when to search a knowledge base, when to call an external API, and when to transfer to a human agent — all through a drag-and-drop visual interface rather than complex prompt engineering.
The Workflows builder is accessible within the ElevenAgents dashboard when configuring a voice agent. It sits between the agent’s voice configuration (which voice, which model) and its system prompt — the Workflow defines the structural logic, and the system prompt fills in the conversational personality and response style. Together they create a complete agent architecture: the workflow defines what to do, the system prompt defines how to sound doing it.
Workflow Node Types
| Node Type | Function | Example Use |
| Start | Entry point of the workflow — where every conversation begins | Greeting message, initial authentication check |
| Message | Agent sends a specific message or response | Fixed response to a specific trigger |
| Branch | Conditional split — different paths based on condition | Verified vs unverified caller, account type, intent detected |
| Tool call | Execute an external function or API call | CRM lookup, calendar check, payment processing |
| Knowledge base search | RAG query against the agent’s knowledge base | Product information lookup, policy retrieval |
| Data collection | Structured collection of information from caller | Name, account number, issue description |
| Handoff | Transfer conversation to human agent or another system | Escalation trigger, specialist routing |
| End | Conversation conclusion — graceful exit | Confirmation message, call summary |
May 2026 Workflow Updates
Flexible Branch Merging
Previously, workflow branches that split at a conditional node had to be resolved before the conversation could continue to downstream nodes. Flexible branch merging (added April 2026) allows branches to merge back into a single path at any point — multiple conditional paths can converge before a shared downstream action. This significantly simplifies workflow design for complex conditional logic where multiple different paths eventually lead to the same outcome.
Built-in Search Tool for RAG
The May 2026 update added a built-in search tool for Retrieval-Augmented Generation directly within the workflow builder — agents can be configured to automatically search their knowledge base when certain conditions are met, without requiring custom tool configuration for basic RAG queries. This makes knowledge base integration into workflows a native drag-and-drop action rather than a custom API call setup.
MCP Tool Support
Workflows now support Model Context Protocol (MCP) tool connections — enabling voice agents to call external tools via MCP servers as workflow nodes. This connects the Workflows builder to the same MCP ecosystem used by Claude Desktop and Cursor, allowing enterprises to reuse existing MCP server infrastructure in their voice agent workflows. A company that has already built MCP servers for internal tools can connect those same servers to ElevenLabs voice agent workflows without rebuilding the integration.
Agent Versioning
Agent versioning allows teams to save and manage multiple versions of a configured agent — including its workflow, system prompt, voice, and LLM configuration — and roll back to previous versions if an update causes unexpected behaviour. For enterprises running production voice agents that handle customer interactions at scale, the ability to version and roll back agent configurations is operationally critical. A bad system prompt update or workflow logic change that degrades agent behaviour can be reversed immediately to a known-good version rather than requiring a debugging cycle under production conditions.
Conversation Topics
Conversation topics (added April 2026) allow agents to track which topics have been discussed in a conversation — enabling downstream workflow logic to reference earlier conversation context. An agent that collected account information in an earlier branch can reference that collected data in a later branch without requiring the caller to repeat themselves. Topic tracking is the mechanism that makes voice agents feel contextually aware rather than amnesiac.
Building a Workflow: Step-by-Step Example
Example: Customer Support Agent with Account Verification
This example builds a basic customer service workflow that verifies callers before providing account information. Step 1: Start node — the agent greets the caller and asks for their account number. Step 2: Data collection node — the agent collects the account number from the caller’s spoken response. Step 3: Tool call node — the agent calls a CRM API with the collected account number to verify the caller and retrieve account status. Step 4: Branch node — conditions: if account verified → path A; if account not found → path B; if account found but flagged → path C. Step 5a (Path A): Knowledge base search node — the agent searches for account-specific information to answer the caller’s question. Step 5b (Path B): Message node — the agent informs the caller their account was not found and offers alternatives. Step 5c (Path C): Handoff node — the agent transfers the call to a human agent due to account flag. Step 6: End node — confirmed resolution and call close.
This workflow encodes complex conditional logic — three different paths based on account verification outcome — that would require intricate prompt engineering to implement reliably without the visual builder. The same logic is maintainable by a non-developer team member through the visual interface.
Workflows vs Prompt-Only Agents: When to Use Each
| Scenario | Prompt-Only Agent | Workflow Agent |
| Simple FAQ chatbot | Sufficient — low complexity | Unnecessary overhead |
| Multi-step customer service | Fragile — hard to maintain | Recommended — visual clarity |
| Conditional access based on caller type | Difficult — prompt bloat | Recommended — branch nodes |
| Multiple API integrations | Complex — coordinate in prompt | Recommended — tool call nodes |
| Human handoff logic | Unreliable — depends on LLM judgement | Recommended — explicit handoff node |
| Rapid prototyping | Fast — prompt is quick to edit | Slower — visual builder setup |
| Enterprise audit requirement | Difficult — prompt not inspectable by non-devs | Recommended — workflow is human-readable |
Three Insights Most Agent Workflow Coverage Misses
1. Workflows Enable Non-Developer Agent Maintenance
The most underappreciated value of Agent Workflows is not technical sophistication — it is that the visual representation makes voice agent logic inspectable and modifiable by non-developers. A contact centre operations manager can review a workflow diagram and understand exactly what the agent does in different conversation scenarios without reading prompt text. A compliance officer can audit the workflow for regulatory alignment without developer involvement. A product manager can propose workflow changes without writing prompt code. This organisational accessibility reduces the developer dependency in enterprise voice agent operations — a significant operational value beyond the technical capabilities of the workflow builder.
2. Agent Versioning Changes the Deployment Risk Profile
Before agent versioning, deploying an update to a production voice agent was a high-stakes operation — a bad update could degrade customer interactions at scale, and recovery required manually reverting changes in the prompt and workflow. Agent versioning changes this risk profile: every deployment creates a new version, and any version can be restored in seconds. This makes continuous improvement of production agents lower-risk — teams can deploy improvements confidently, knowing that regression to a known-good state is a one-click operation rather than an emergency debugging session.
3. MCP Integration Makes ElevenLabs Agents Part of the Broader AI Infrastructure Ecosystem
The addition of MCP tool support in Workflows is strategically significant beyond its immediate functionality. MCP is becoming the standard protocol for AI tool integration across the industry — Claude, Cursor, Windsurf, and a growing list of AI tools support MCP servers. By making ElevenLabs voice agents MCP-compatible tool consumers, ElevenLabs integrates its agents into the same tool ecosystem as other AI systems in an enterprise’s environment. A company’s internal MCP servers for CRM access, calendar integration, or data retrieval can serve both Claude-based AI applications and ElevenLabs voice agents through the same interface — reducing the total integration burden for enterprises running multiple AI systems.
Agent Workflows in 2027
The Workflows builder will evolve toward more sophisticated conditional logic — multi-condition branches, percentage-based traffic splitting for A/B testing conversation strategies, and time-based conditions (different workflow paths during business hours versus after hours). Agent analytics integrated with the workflow visualisation — showing which conversation paths are most frequently taken, where callers drop off, and which tool calls most often fail — will turn the workflow diagram into a live operations dashboard. And workflow templates from ElevenLabs and the creator community will provide starting points for common agent use cases, reducing the setup time for new deployments.
Key Takeaways
- ElevenLabs Agent Workflows is a visual no-code builder for voice agent conversation logic — branches, tool calls, data collection, knowledge base search, and handoff nodes in a drag-and-drop interface.
- May 2026 updates added: flexible branch merging, built-in RAG search tool, MCP tool support, agent versioning, and conversation topic tracking.
- Use Workflows for complex agents with conditional logic, multiple integrations, and human handoff requirements. Use prompt-only for simple FAQ agents.
- Agent versioning reduces deployment risk — every update creates a rollbackable version.
- Workflows make agent logic inspectable by non-developers — enabling compliance auditing and business-side maintenance without developer involvement.
Conclusion
ElevenLabs Agent Workflows is the infrastructure that transforms voice agents from developer-maintained prompt artefacts into enterprise-grade operational systems that business teams can inspect, audit, and maintain. The May 2026 updates — MCP tool support, flexible branch merging, agent versioning, built-in RAG search — significantly advance the platform toward production-ready enterprise deployment. For businesses building voice agents that handle real customer interactions at scale, the investment in Workflow-based architecture rather than prompt-only design pays back in maintainability, auditability, and the reduced developer dependency that makes ongoing agent improvement operationally sustainable.
Frequently Asked Questions
What are ElevenLabs Agent Workflows?
A visual, no-code conversation logic builder within ElevenAgents. Teams can define the conditional paths, tool calls, data collection, and handoff logic of a voice agent through a drag-and-drop interface rather than encoding all logic in a system prompt.
What was added to Agent Workflows in May 2026?
The May 2026 release added flexible branch merging, built-in knowledge base search tool for RAG, MCP tool support (connect external MCP servers as workflow tools), agent versioning (rollback to previous configurations), conversation topic tracking, and enhanced pronunciation rules.
Do I need to code to use ElevenLabs Agent Workflows?
No — the Workflows builder is a visual drag-and-drop interface accessible to non-developers. Tool call nodes that connect to external APIs may require developer setup of the underlying API endpoints, but the workflow logic itself is built visually.
What is agent versioning in ElevenLabs?
A feature that saves each agent configuration as a versioned snapshot — including workflow, system prompt, voice, and LLM settings. Any version can be restored in seconds, reducing the risk of production deployments and enabling rapid rollback if an update degrades agent behaviour.
How does MCP tool support work in Agent Workflows?
MCP server connections can be added as tool call nodes within the workflow builder. External tools exposed through MCP servers — CRM systems, calendar services, internal databases — are callable from the agent workflow without custom API integration code, using the same MCP server infrastructure that connects to Claude Desktop and Cursor.
Methodology
Agent Workflows features from ElevenLabs official ElevenAgents documentation and workflow builder product page. May 2026 updates from Releasebot ElevenLabs May 2026 release notes. MCP tool support from ElevenLabs changelog (April 2026) — MCP pre_tool_speech and response_timeout_secs. Agent versioning from ElevenLabs changelog (January 2026). Flexible branch merging from ElevenLabs changelog (April 13, 2026). Conversation topics from ElevenLabs changelog (April 13, 2026). This article was drafted with AI assistance and reviewed by the editorial team at ElevenLabsMagazine.com.
References
ElevenLabs. (2026). ElevenAgents documentation. https://elevenlabs.io/docs/eleven-agents
Releasebot. (May 2026). ElevenLabs May 2026 release notes. https://releasebot.io/updates/eleven-labs
ElevenLabs. (2026). Changelog — workflows and agent updates. https://elevenlabs.io/docs/changelog
