Agentic AI Platform Change Log

Deployment / data residency

Vapi can now split an assistant's live traffic across published versions by percentage when publishing, so teams can canary a new version and roll back instantly, with a live traffic view in the assistant header. Call logs can now be filtered by assistant version.

Prompt or tool changes to a production voice agent no longer have to reach every caller at once, which lowers the risk of a bad release and gives a measured comparison on real calls.

MCP / tool calling / API

Agent Phone, now in public beta from Vapi, is a remote MCP server that lets assistants such as Claude Code and Cursor place phone calls for a user. After a one time approval in the browser, the agent dials, works through phone menus, waits on hold, talks to the person and reports back, with no telephony setup.

Any MCP capable agent can now finish tasks that need a live phone call, which extends agent reach to businesses that only work by phone. The revocable 30 day credential defines how long the agent may call on its own.

MCP / tool calling / API

Vapi added a serviceTier setting that lets assistants request faster OpenAI processing for supported GPT-5.6 and GPT-5.5 models. Its September 21 update also adds provider request IDs to call logs for language model, transcription, and voice services when customers bring their own keys.

Voice teams can test a latency and cost tradeoff without changing their agent workflow, while request IDs make provider support investigations easier to correlate with specific calls. Fast processing uses the provider's corresponding rates and may be downgraded under load.

Observability / auditability

Vapi's Composer can now generate simulation scenarios, tester personalities and full test suites automatically, and assistants can fall back to another Anthropic model on Bedrock from the same provider.

Automated test generation lowers the effort of testing a voice agent before launch. Same provider fallback keeps a call alive during a model outage without changing data residency.

Agent capability

Vapi introduced native outbound calling campaigns, allowing users to upload contacts, select an assistant, and track results directly from the dashboard. The update also adds a prompt editor with versioning snapshots, a new webhook for processing call artifacts as they upload, and Azure region pinning for GPT-5 models.

Native campaign management eliminates the need for third-party dialing orchestration, simplifying outbound operations. Azure region pinning also allows enterprise buyers to enforce data residency and reduce latency, while the new prompt editor and versioning tools improve the developer experience for maintaining agent configurations.

Workflow orchestration

Vapi introduced Conditional Structured Outputs, allowing structured data to be generated only when specific conditions are met during a call. The update also adds support for Deepgram's Aura-2 German voices to the platform.

Developers can enforce stricter logic on data extraction, reducing unnecessary processing and improving the accuracy of call logs. The new German voices also expand the platform's multilingual capabilities for European deployments.

Agent capability

Vapi introduced Model Intelligence, allowing users to configure an assistant's transcriber, LLM, and voice models in a single click using pre-tuned presets like Balanced, High Intelligence, Ultra Fast, or Cost Saver. The update surfaces latency, cost, and quality metrics directly in the dashboard. End-of-call reports also now feature short-lived presigned URLs for secure recording downloads and are reliably delivered even when using Zero Data Retention.

The model presets and visible metrics help teams deploy appropriate model combinations that balance cost and performance without running manual benchmarks. The presigned URLs and robust zero data retention support ensure that compliance-focused engineering teams can safely and programmatically access post-call artifacts.