From prototype to production: the thunk lifecycle
A thunk is an application, and like any application it has a life after it is first built: it is tested, reviewed, promoted into production, watched, tuned, and changed over time. This article walks through that lifecycle and points to the article that covers each stage.
1. Design
You describe the process in the Workflow Plan, give the AI agents the documents they need in Content Folders, and connect the tools they use in Connections and Custom-built tools. See Thunk Application Model.
2. Test
Before anyone relies on a thunk, check that it produces the right results — and keep checking as it changes.
Manual testing — drive work items through the workflow yourself and watch what the AI agent does. See Manual testing.
Automated tests — turn representative work items into test work items with assertions, and re-run them after every change. See Automated tests for workflow steps.
Tool tests — test each custom tool on its own. See Automated tests for custom tools.
Test plan design and Production readiness and quality control describe what a good test plan covers.
3. Review
A design review checks the thunk's design for common problems — unclear instructions, missing inputs, tools without tests — before it is promoted. Run it from the Review tab of the Deploy pane. See Workflow Review.
4. Version
Every change to a thunk's definition — by you, a teammate, or the AI agent — is recorded automatically, so you can see what changed, when, and who changed it. See Versioning Support for your Thunk.
5. Promote
A thunk moves through three phases — Prototype, Testing, and Production — from the Deploy pane. Each phase runs against an environment your administrator sets up, which controls the AI models and tools the thunk can use. Promoting forward asks you to confirm the thunk has been reviewed. See Lifecycle of a Thunk. The first tab of the Deploy pane, Pause / Restart, stops and resumes all work on the thunk; see Pause and restart a thunk.
6. Monitor
Once a thunk is doing real work, the Monitor pane (for thunk admins) shows how it is doing. See The Monitor pane: what each tab is for.
Tab | What it tells you | Article |
Errors | Errors users have run into while the thunk was running | |
Cost | What the thunk spends on AI, per day and per work item | |
Time | How long work items take, and which ones are slow | |
Reports | Save filtered views of your work items as reports | |
Tracking | Send events about the thunk's work to your own monitoring or analytics platform | |
Log | The work happening behind the scenes to keep the thunk running |
To check a running thunk across all of these — and the order to check them in — see Checking a thunk's health: where to start.
7. Diagnose and tune
When a work item goes wrong or takes too long, open it and look at what the AI agent actually did. Why so long? breaks a run's time into model time, tool time, and the rest; What went wrong? explains a failed run. See Analysis of AI behavior and AI reliability troubleshooting. For what drives cost and speed in general, see Understanding LLM costs and latency.
After any fix, re-run your automated tests before promoting the change.
8. Change AI models
AI models improve and are eventually retired, so a long-lived thunk will change models at some point. Test the new model on your test work items, compare results, cost, and time, then switch. See Move a thunk to a new AI model and Supported AI models and what happens when they change.
9. Upgrade
When a thunk was built on an older version of the platform, thunk admins see an Upgrade to V… button at the top of the thunk's left navigation. It brings the thunk up to the current platform version. After upgrading, re-run your automated tests to confirm the thunk still behaves as expected.