ADK explained: how PolyAI lets any team build and deploy dialog agents

How developers build and ship PolyAI agents from their own environment — git, IDE, CI/CD, no lock-in.

Brady Walker Senior Content Marketing Manager
4 min
Share

Most AI platforms give developers one option: learn the platform's proprietary way of doing things. PolyAI's Agent Development Kit takes the opposite position. ADK brings the platform to the developer's existing environment — git workflow, IDE, CI/CD pipeline — and leaves it there.

Here's what that means in practice.

ADK in one paragraph

ADK is a CLI and Python package that lets engineers build, version, and deploy PolyAI dialog agents from their own development environment. It connects to Agent Studio — PolyAI's no-code build platform — via poly push, syncing local changes without requiring manual UI work. All execution happens inside Agent Studio; ADK is the local authoring layer. Think of it as git for Agent Studio: branch-based development, structured version control, and full CI/CD automation via the Agents API.

Two audiences, one platform

The reason ADK matters for enterprise deployments isn't the developer experience in isolation. It's that ADK and Agent Studio's no-code builder operate on the same platform and under the same governance model.

A customer service manager can build an FAQ flow in Agent Studio without touching code. A developer on the same team branches that project in ADK, writes a Function step that calls a backend API, and pushes the change through sandbox, pre-release, and live via an automated pipeline. Both work in the same environment. Both changes go through the same versioning, approval gates, and deployment controls.

The alternative — a developer tool for engineers and a separate no-code tool for business teams — means two governance models, two sets of access controls, and eventual configuration drift between what engineering shipped and what the business team thinks is live. ADK eliminates that split.

What you can build with PolyAI’s ADK

ADK gives engineering teams native workflow support for any agent type on the PolyAI platform:

  • Account management — authenticate callers, retrieve account data via Function steps, route based on account status
  • Call routing and triage — intent classification at the front of a call, conditional handoff logic to the right queue or live agent
  • FAQ and self-service — knowledge retrieval via Connected Knowledge sources (Zendesk, Gladly, or file upload), managed topics for precise response control
  • Transactional flows — payment handling, booking, and form-based data collection with entity extraction and backend API calls

All of these deploy across voice, web chat, SMS, WhatsApp and in-app from the same agent configuration. Build once, run across channels.

How it fits into enterprise engineering workflows

ADK is designed for teams that can't afford a bespoke deployment process for every agent change. The workflow has three steps:

  1. Build locally. Write agent configuration in your IDE using ADK's CLI and Python package. Work in branches like any other codebase — multiple engineers can develop separate agent features simultaneously without conflicts.
  2. Push to Agent Studio. poly push syncs local changes to Agent Studio's sandbox environment for testing, without manual UI steps. The change is live in sandbox immediately and doesn't touch pre-release or production until you say so.
  3. Promote to production. The Agents API exposes publish, promote, and rollback endpoints. The full pipeline — sandbox to pre-release to live — runs automated. Rollback is available at each environment independently, so a production issue doesn't require touching other environments to fix.

Independent deployments with shared governance

For organizations running multiple deployments across business units or regions, Agent Studio lets each team release changes to their own agent variant independently. A regional team pushing an update to their flow doesn't trigger a regression test across every other team's deployment. Each domain is isolated — its own prompts, facts, and business rules — but governed under the same account-level access controls. ADK users can run independent CI/CD per domain, with rollback available at each environment independently.

The access control model is the same across ADK and the no-code builder: role-based permissions, quarterly access reviews, and a full audit trail of every change. A developer pushing via ADK and a business user editing in the UI operate under identical governance.

Getting started in minutes

curl -LsSf https://astral.sh/uv/install.sh | sh # install uv (skip if you have it)
uv venv --python=3.14 --seed
source .venv/bin/activate
pip install polyai-adk
poly start # self-serve sign-up, API key, and first project in one go

Resources

See all resources