Intel / article

Unpacking Elevenagents Orchestration Engine

Source: elevenlabs.io
Author: ElevenLabs editorial sitemap
Date: Published Sep 14, 2026

Why it matters

Shows how a production voice agent splits work across models, how conversation history is shared, and when to choose a single agent over a multi-agent workflow. Useful for designing low-latency agent pipelines.

Key takeaways · AI-distilled

ElevenLabs says its orchestrator predicts when a user has finished speaking to cut perceived latency, so one conversational turn can trigger several LLM generation requests with the same .
Every enabled tool adds its name, description and parameter schema to the prompt. ElevenLabs advises that the tool description say what the tool does and returns, while the says when to call it, so tool definitions stay reusable across agents.
Tools run in three modes: Immediate (default, optionally with filler speech while it runs), Post-Tool Speech (waits until the agent finishes talking, for transfers or payments the user can still interrupt) and Async (background work whose result the reply does not need).
In Workflows, LLM-evaluated transition conditions run in parallel outside the active 's prompt, and the orchestrator is rebuilt at every transition so each sub-agent sees only its own tools and knowledge. The post presents this as trading flexibility for .
Under Zero Retention Mode, call data stays in memory and the LLM choice is limited to providers contractually barred from retaining or training on it, which the post lists as Gemini and Claude. Other models need the customer's own agreement plus Safety team approval.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
subagent — A separate agent spawned by another to do one scoped piece of work in its own context, returning only the result.
determinism — Whether the same input reliably produces the same output — something LLM systems mostly lack, which changes how you test and debug them.

Read the source elevenlabs.io