Vapi launches Agent Loop to help voice agents improve after every call

A continuous improvement loop for voice agents

Agent Loop sits on top of Vapi's platform and automates a cycle that the company says many customers have been assembling themselves. It creates an agent, scores calls against a selected goal, identifies missed outcomes, drafts a fix, tests the change, sends it to part of live traffic and measures the result.

The process can then repeat through what Vapi calls recursive self-improvement. The system can adjust models, architectures, voices, configurations and tools, while the underlying Vapi functionality remains available in the dashboard. Vapi says each step is explainable and configurable rather than a black box.

Users remain in control of approvals. Decisions arrive in an inbox connected to Slack, and Vapi says nothing ships without a person saying yes. As users gain confidence in the agent's judgment, they can allow it to approve more changes independently.

From business goal to tested deployment

Users begin by defining a goal, such as resolution rate, and provide Composer, the AI agent built into the Vapi dashboard, with relevant context. That can include process steps, integrations, knowledge bases and real call recordings. Composer sets up tools and prompts, selects models and architecture, and assembles an initial version.

As calls arrive, Agent Loop reviews them against the goal, examines misses and groups them into issues. Composer then prepares an improvement for the most important issue, including evidence calls and the exact prompt or configuration difference. The proposed version runs through hundreds of simulated conversations before deployment.

If the change performs well, it can be rolled out more broadly. If the measured number falls, Vapi says the system rolls the change back automatically and records what it learned. Every Agent Loop agent remains a real Vapi assistant, so a customer experience lead can manage it through Agent Loop while an engineer retains access to the dashboard and API for the same assistant.

Early access features and dashboard plans

Vapi says Agent Loop is already being used by customers, with general availability planned for early next year. The company also plans to bring its underlying primitives to the dashboard over the coming months, whether customers use the full self-improvement workflow or incorporate individual tools into existing processes.

Simulations are available in the dashboard now and use AI callers with distinct personalities to run hundreds of conversations against a proposed change. Traffic Splitting is also available now, allowing teams to send part of their live calls to a new version before a broader rollout.

Goals, Call Health and Experiments are coming to the dashboard soon. Goals define pass or fail conditions in business terms. Call Health analyzes calls, ranks recurring problems and provides transcript evidence plus proposed prompt or configuration changes. Experiments compare versions against measures such as a goal, issue rate, cost or latency, and can help select a winner and shift traffic.

For Vapi users, the launch brings a structured way to connect monitoring, testing, experimentation and deployment. The company says customers can use the complete loop or retain deeper control over its individual components.