What happened
On October 29, 2026, Cartesia announced Managed Agents, which it describes as an end-to-end platform for enterprise voice conversations. The product combines Cartesia's Ink speech-to-text and Sonic text-to-speech models with a language model selected by the developer. It packages listening, turn-taking, reasoning, tool calling, and speech as a single system rather than asking a team to independently assemble those layers.
The platform includes telephony, knowledge bases, and tool integrations. Cartesia positions it for customer support, appointment scheduling, sales and lead qualification, recruiting and screening, and collections, with industry examples spanning finance, retail, hospitality, and healthcare. It also says Managed Agents has already powered tens of millions of conversations. The launch material describes support for 44 languages, including the ability to switch among five languages during a call at launch. Cartesia lists agent pricing at $0.03 per minute, plus $0.014 per minute for calls using a Cartesia phone number.
Cartesia's product argument is architectural. It says conventional cascaded voice pipelines can lose conversational quality when their parts are not carefully co-designed, particularly around interruptions and turn-taking. Conversely, it characterizes speech-to-speech systems as natural sounding but harder to control for enterprise tasks. Managed Agents uses a cascaded speech-to-text, LLM, and text-to-speech setup, but Cartesia says the voice models and harness were designed together to preserve both conversational dynamics and the controllability needed for tasks involving tools.
The company supports that claim with two benchmarks. On Artificial Analysis's Speech Agent Arena, which uses blind human preferences on tasks including appointment booking and takeout ordering, Cartesia reports a first-place ranking among 30 agents with a preference Elo of 1115 at launch. The listed evaluation uses GPT-5.5 as the reasoning model. On ServiceNow's EVA-Bench, a benchmark covering more than 200 enterprise CSM, ITSM, and HR scenarios, Cartesia reports 87.4% accuracy and 90.7% experience. The published comparison places Managed Agents ahead of the listed systems on both measures, though these are benchmark results rather than evidence of performance in every deployment.
What the archive adds
This is a new product event, but it extends the voice-agent product boundary described in Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages. That note focused on packaging speech models and a real-time SDK into a deployable workflow; Cartesia adds a managed enterprise layer with telephony, knowledge, integrations, and per-minute operation.
Trend signal
The launch supports Harness architecture outlives model choice. Cartesia lets the developer choose the LLM while making the surrounding system—turn-taking, speech models, tools, knowledge, and telephony—the managed product. Its benchmark claim is specifically an argument that the orchestration layer, not only the model identity, determines practical voice-agent quality.
Connections to other events
The earlier Voxtral 3 note packaged speech models with a real-time agent SDK for deployment. Cartesia takes a different delivery route: instead of presenting primarily an SDK and models, it offers a managed service with the operational pieces required to put an agent on the phone. Both events make deployment architecture part of the voice product rather than treating speech generation as an isolated model call.
Why it matters
For a team building a phone-based agent, the expensive work is usually not only selecting a voice model. It is coordinating transcription, response generation, speech, barge-in behavior, tool calls, retrieval, phone infrastructure, observability, and failure handling. A managed platform can reduce that integration work, while the developer-selected LLM preserves a degree of control over reasoning and provider choice.
The more consequential claim is that a cascaded system can score well on both task accuracy and conversational experience when its parts are designed as one harness. Teams should therefore avoid making an architecture decision solely on whether a provider offers a native audio model. A native speech-to-speech design may be appropriate for a particular interaction, but a controllable cascade with strong turn-taking and tool orchestration can be a serious competing option for enterprise workflows.
What happens next
- Voice-agent evaluations will increasingly compare complete configured systems rather than only speech models. (6 months, 70%). Watch for: benchmark submissions that disclose the reasoning model, tool configuration, and telephony or harness setup.
- Managed voice platforms will compete on operational controls as well as voice quality. (12 months, 65%). Watch for: additions around evaluation, governance, integrations, and deployment controls.
Limits and uncertainty
The source is Cartesia's launch announcement, so its product positioning, scale claim, pricing presentation, and interpretation of the results should be treated as vendor statements. The benchmark figures are useful signals, but they do not establish performance for a specific company's data, calling geography, tool reliability, latency target, compliance requirements, or conversation design. The Speech Agent Arena result also uses GPT-5.5 as the reasoning model, meaning it is not a measurement of the speech layer in isolation. The material does not provide a full cost model for every deployment configuration or a direct operational comparison with alternatives.
Skill to build now
Create a voice-agent evaluation script before choosing a stack. Include representative calls with interruption, ambiguity, tool failure, knowledge retrieval, escalation, and language switching; score both task completion and the caller's experience. Run the same script across at least one managed cascade and one alternative architecture, while recording the reasoning model, prompts, tools, latency, and per-minute cost. That makes the architecture decision reproducible rather than a reaction to a leaderboard rank.
Reader takeaway
Managed Agents is not merely a new voice model endpoint. It is Cartesia's attempt to make the entire voice-agent harness—speech, turn-taking, tools, knowledge, and telephony—the product. For builders, the important test is whether that managed boundary improves verified task completion and caller experience in their own workflow at an acceptable operational cost.