workflow the agent does not follow it | ElevenLabs | Page 1

solar basinBOT Dec 26, 2025, 5:57 PM

#

AI-generated Answer for sari_33713

It sounds like you’re using Agent Workflows and the live agent isn’t taking the path you designed. Here are the specific checks and fixes documented by ElevenLabs:

Verify edges and conditions
- Ensure every node is connected by a forward edge and that each edge has the correct type (LLM Condition, Expression, or None). LLM Conditions are evaluated at runtime to choose the next node. (elevenlabs.io)
Use a Dispatch tool node when a tool must run
- Tools inside a Subagent are optional; a Dispatch tool node is a dedicated execution point that guarantees the tool is called and lets you route on success/failure. If a tool call is critical to the path, use a Dispatch node instead of relying on general tool availability. (elevenlabs.io)
Check transfer configuration (if handoffs aren’t happening)
- Agent-to-agent: add the transfer_to_agent system tool and define transfer rules (target agent + natural-language condition). The LLM triggers transfers based on these rules. (elevenlabs.io)
- Human transfer: for phone calls, enable transfer_to_number and configure the destination and messages. Chat widgets don’t support this tool. (elevenlabs.io)
Confirm you’re running the correct version/branch
- If Versioning is enabled, edits are committed as immutable versions on branches and traffic can be split across branches. Drafts are per-user and not live. Make sure the branch with your updated workflow is receiving traffic (or include_draft when testing via API). (elevenlabs.io)
Add automated tests to reproduce the misroute
- Use Agent Testing to simulate conversations and assert that specific tools were called and specific workflow paths were taken. Run from the Tests tab or via the CLI (elevenlabs agents test <agent_id>). (elevenlabs.io)
Inspect results to pinpoint failures
- Use Conversation Analysis to evaluate outcomes and extract structured signals that help you see where routing diverged (e.g., condition not met vs. tool failure). (elevenlabs.io)

If you can share which node isn’t reached and whether you’re using transfers or tool calls, I can point to the exact configuration section to adjust. (elevenlabs.io)

stray wraith Dec 26, 2025, 5:59 PM

#

what is a dispatch node? a subagent?

solar basinBOT Dec 26, 2025, 5:59 PM

#

AI-generated Answer for sari_33713

Dispatch node: A “Dispatch tool node” is a workflow node that executes a specific tool call at that point in the conversation and guarantees the tool is invoked. It also lets you route the flow on success vs. failure of the tool call (separate edges for each outcome). (elevenlabs.io)
Subagent: A “Subagent node” lets you change or override the agent’s configuration for a segment of the flow—e.g., adjust or replace the system prompt, switch LLM, change voice settings, and update available tools or attached knowledge base items. Changes apply on top of, or in place of, the base agent for that node. (elevenlabs.io)

Sources:
Workflows | ElevenLabs Documentation
Workflows | ElevenLabs Documentation

stray wraith Dec 26, 2025, 6:01 PM

#

i dont see a dispatch option only subagent, agent transfer, phione number transfger, tool, and end

solar basinBOT Dec 26, 2025, 6:01 PM

#

AI-generated Answer for sari_33713

You’re not missing it—“Dispatch” is the Tool node. In the Workflows docs it’s called “Dispatch tool node,” but in the UI it appears simply as “Tool.” It’s used to force a specific tool call at that point in the flow and lets you branch on Success vs Failure from that tool execution. (elevenlabs.io)

Quick distinctions:

Tool (Dispatch): guarantees a tool call and supports success/failure routing. (elevenlabs.io)
Subagent: changes the agent’s config/KB/tools for that segment but doesn’t force a tool call. (elevenlabs.io)
Agent transfer and Transfer to number: hand off to another AI agent or to a phone number (human). (elevenlabs.io)
End: terminates the conversation. (elevenlabs.io)

If you want step-by-step on wiring the success/failure edges for the Tool node, I can walk you through it. (elevenlabs.io)

stray wraith Dec 26, 2025, 6:02 PM

#

thats not the issue i have very specific instructinos on each node but teh agent just does not follow them consistently at all

solar basinBOT Dec 26, 2025, 6:02 PM

#

AI-generated Answer for sari_33713

Understood. Inconsistent node behavior usually comes from non‑deterministic routing or prompt scope bleed. These steps tighten determinism and prompt control:

In each Subagent node, replace (don’t append) the base prompt when you need strict behavior, and scope the node’s own Tools and Knowledge Base there. Subagent nodes can append or override system instructions and update tool/KB availability per node. (elevenlabs.io)
For must‑run actions, use a Tool node (the “Dispatch tool node”) so the tool is guaranteed to execute, and branch on its Success/Failure edges instead of leaving the choice to the LLM. This is the recommended way to make flows deterministic (e.g., auth) rather than model‑decided. (elevenlabs.io)
Prefer Expression‑based edges driven by tool outputs/dynamic variables for critical routing; avoid pure LLM Condition edges for policy gates. The workflows system supports Expression edges, and the auth guide recommends routing from tool results or “transfer expressions.” (elevenlabs.io)
If the agent ignores tool‑usage instructions, make tool calls explicit in the system prompt and fully document tool parameters; this improves tool‑calling reliability. See the prompting guide sections on “Tool configuration for reliability.” (elevenlabs.io)
Structure each node prompt for reliability: separate sections (Goal, Steps, Tools, Guardrails), keep instructions concise, and emphasize critical rules. This reduces instruction bleed and improves compliance. (elevenlabs.io)
Use Subagent nodes to switch LLMs on hard decision points (trade latency vs. tool‑calling vs. reasoning). The prompting guide lists model tradeoffs and recommendations; Subagents can change the model per node. (elevenlabs.io)
Turn failing conversations into regression tests (Agent Testing) so the platform automatically checks the expected behavior on each change. (elevenlabs.io)
Stabilize deployments with Versioning: freeze a known‑good configuration, test changes on a branch, and roll out by traffic split to avoid “surprise” behavior shifts. (elevenlabs.io)
Debug symptoms in Call History: look for errors like “Missing dynamic variable after agent transfer” or “Override is not allowed for this AI agent,” which indicate config, variable, or override issues rather than prompt quality. (help.elevenlabs.io)

If you can share one conversation ID

stray wraith Dec 26, 2025, 6:08 PM

#

conv_7001kddw6h85e8vs779gwjv1dty1

#workflow the agent does not follow it