Twilio in 2026: ConversationRelay, Agent Connect and the conversation layer. The phone side of voice agents, explained for buyers

Every AI phone agent you have heard about has a carrier underneath it, and more often than not that carrier is Twilio. In 2026 Twilio stopped being only the pipe and started shipping the pieces that sit between the phone line and the model. This note explains what those pieces are, which ones a UK business should care about, and where Twilio ends and your agent begins.

The 2026 line-up

At SIGNAL San Francisco on 6 May 2026, Twilio launched four capabilities into general availability and upgraded a fifth:

• ConversationRelay. The real-time bridge for voice AI: Twilio handles the call audio, speech recognition, text-to-speech, interruption handling and streaming, and hands your application text events over a WebSocket. You bring whichever language model you like. In 2026 it added PCI compliance, HIPAA eligibility, Insights, and support for Deepgram Flux, a speech model that merges transcription and end-of-turn detection. Twilio says Flux cuts agent response latency by 200 to 600 milliseconds and false interruptions by about 30 percent, and it exposes tuning knobs such as eotThreshold and ignoreBackchannel so "yeah" and "uh-huh" stop cutting the agent off
• Agent Connect. An open-source, self-hosted framework that connects your AI agent to Twilio's voice and messaging channels and handles real-time streaming, session and identity management. It is model-agnostic, so you can change AI provider without reworking the Twilio integration
• Conversation Memory. Persistent context across channels and sessions, so the agent on the phone knows what happened on WhatsApp yesterday
• Conversation Orchestrator. Routing and hand-off between AI agents, human agents and channels
• Conversation Intelligence. Analysis of voice and messaging as it happens: intent, sentiment, outcomes, guidance to human agents, and triggers for automated workflows. It integrates natively with ConversationRelay, so you can review how your AI agent performed on real calls

Twilio also rebuilt its Console with a developer Workbench, and SIGNAL São Paulo on 6 August added Interconnect for private, dedicated network connections.

Why this matters more than it sounds

The hard problems in a phone agent are rarely the model. They are: knowing when the caller has finished speaking, not talking over them, keeping latency low enough that the pause feels human, handling a transfer, and being able to prove afterwards what was said. ConversationRelay plus Flux addresses the first three. Orchestrator addresses the transfer. Intelligence and Insights address the proof.

That is why all-in-one platforms run on Twilio too. The question for a buyer is whether to use Twilio directly with a framework such as Pipecat and a model such as Grok Voice 3.0, or to pay a platform to assemble it. We work through that in build or buy a phone agent.

Who should care, who should wait

Use Twilio directly if

• You already run Twilio numbers, SIP trunks or messaging and want the agent on the same account, billing and compliance footprint
• You need HIPAA eligibility or PCI scope on the voice path and want it from the carrier rather than a reseller
• You have, or are hiring, an engineer to own Agent Connect or a Pipecat pipeline
• You want to choose and change the model, STT and TTS vendors yourself

Go through a platform if

• Nobody will own the integration code
• The use case is a standard front door (answer, book, escalate) and a platform already does it well for your sector
• You need a dashboard a receptionist can use on day one

UK-specific checks

• Numbers and regulation. UK geographic and non-geographic numbers, Ofcom-compliant caller ID, and the recording notice at the start of the call are your responsibility, not the model's
• Data residency. Ask where ConversationRelay processes audio and where Intelligence stores transcripts for UK customers, and get it in writing for your DPIA
• Costs. Twilio bills per minute for voice, separately for ConversationRelay, and separately again for STT and TTS partners. Model cost (for example $0.08 a minute for Grok Voice or Eleven Realtime) is on top. Build a per-call cost before a per-month budget
• Escalation. Test the hand-off to a human with Orchestrator under load. A failed transfer is the complaint you will hear about

Twilio vs the alternatives

Bottom line

Twilio in 2026 is a credible place to build a phone agent directly, not just the number you rent for someone else's platform. ConversationRelay with Flux solves the turn-taking problem that makes most phone agents feel robotic, Agent Connect gives you an open-source bridge that does not lock you to a model, and Intelligence gives you the evidence trail regulated UK businesses need. The price is owning the integration. If you have the engineer, it is a good trade.

Related reading

• Build or buy a phone agent in 2026
• Pipecat in 2026: from 1.0 to 1.15
• Voice AI for the front door
• The calls you miss are the revenue you lose
• Twilio on the marketplace · Browse customer service tools

Sources: Everything Twilio launched at SIGNAL 2026 · ConversationRelay now supports Deepgram Flux (changelog, 6 May 2026) · Integrate Deepgram Flux with ConversationRelay · SIGNAL São Paulo press release · UC Today on SIGNAL 2026

Erhan Timur, Founder, Digital by Default