PolyAI's ADK lets engineers build voice agents in their own coding tools

What ADK actually is

PolyAI builds AI agents that hold conversations with customers, such as on a phone line. In an explainer, poly.ai described its Agent Development Kit, or ADK, as a way for engineers to build these agents without leaving the tools they already use.

Technically, ADK is a command-line tool (software you control by typing commands) plus a Python package (a bundle of code). Engineers use it to write, track and release PolyAI agents from their own computers. PolyAI compares it to git, the standard system programmers use to track changes to code, but aimed at its Agent Studio platform.

Agent Studio is PolyAI's no-code builder, and it is where agents actually run. ADK is only the local place where the work gets written.

One platform, two kinds of builders

PolyAI's main argument is about teamwork, not developer convenience. A customer service manager can build an FAQ flow in Agent Studio without writing code. A developer on the same team can then branch that project (make a separate working copy) in ADK and add a Function step, which is a piece of logic that calls a company's backend system.

Both people's changes go through the same versioning, approval gates and release controls. PolyAI says the alternative is a developer tool for engineers and a separate no-code tool for business staff. That means two sets of rules, two sets of access controls, and a growing gap between what engineering shipped and what the business thinks is live.

From laptop to live call

The workflow has three steps.

• Build locally. Engineers write the agent's configuration in their code editor and work in branches, so several people can build different features at once without colliding.
• Push to Agent Studio. The command poly push sends the changes to a sandbox, a safe test area. Nothing reaches pre-release or production until someone chooses to promote it.
• Promote to production. An interface called the Agents API (a way for programs to talk to each other) offers publish, promote and rollback actions. The route from sandbox to pre-release to live can run automatically.

Each environment can be rolled back on its own. A problem in production can be fixed without disturbing the test stages.

What teams can build with it

PolyAI lists four kinds of agent that ADK supports. Account management agents verify callers, pull up account data and route them based on status. Call routing agents work out what a caller wants at the start of a call and hand them to the right queue or a live person.

FAQ agents draw on connected knowledge sources such as Zendesk, Gladly or an uploaded file, with managed topics to control exactly what gets said. Transactional agents handle payments, bookings and forms, collecting details and calling backend systems.

The same agent setup can serve phone calls, website chat, text messages, WhatsApp and in-app conversations, so a team does not rebuild it for every channel.

Many teams, shared oversight

Large organizations often run several deployments across business units or regions. PolyAI says Agent Studio lets each team release changes to its own agent variant independently. A regional update does not force a round of regression tests (re-checks that old features still work) across everyone else's deployments.

Each domain keeps its own prompts, facts and business rules, yet all sit under account-level access controls. Teams using ADK can run a separate automated release pipeline for each domain.

The rules are the same whether someone uses ADK or the no-code builder: permissions tied to roles, regular access reviews, and a complete record of every change. A developer pushing code and a manager editing in the browser are held to the same standard.

What to keep in mind

This is PolyAI describing its own product, so the claims about smoother governance come from the vendor. The explainer does not give pricing, and our newsletter noted that price was not stated.

The phrase worth noting is that teams of different skill levels can work in one system. PolyAI also announced a Builder MCP for working inside code editors, which fits the same direction: make building conversational agents less of a specialist job.