FlakeWarden uses two narrowly-scoped agents, both authored in UiPath Agent Builder and orchestrated by Maestro. They follow Taqi Jaffri's single-responsibility / "small in scope, big in impact" principle: each agent does one thing, is grounded in trusted data, and is gated by a human.
| Agent | Responsibility | Grounding (context grounding / RAG) | Output |
|---|---|---|---|
| Triage Classifier | Classify an ambiguous failure as real_defect / flaky / environment | Test Manager stack traces, DOM/selector diffs, commit messages, runner logs | structured verdict + confidence + (flaky only) proposed fix |
| Repair Agent | Draft a selector/synchronization repair for a confirmed flaky test | The test's object repository / selectors + the proposed fix | a draft PR-style patch sent to Action Center |
Naming note: the Repair Agent is FlakeWarden's own Agent Builder agent that drafts a fix for human review. It is distinct from UiPath's GA Healing Agent™ (a platform feature that performs runtime self-healing of UI automations). After a reviewer approves the draft, FlakeWarden can hand off to UiPath's Healing Agent to apply it. We do not claim to have built UiPath's Healing Agent.
The Triage Classifier is invoked only on failures the deterministic flake-scorer could not confidently bucket, so the model is spent only where it adds value.
classifier_prompt.txt— the system prompt. The localAnthropicClassifierand the UiPath Agent Builder agent use the same prompt text, so behaviour is identical between local eval and the deployed agent.classifier_agent.json— Agent Builder definition (model, inputs, grounded sources, output schema, eval set, guardrails).healing_agent.json— Repair Agent definition.
Per Taqi Jaffri's evaluation-driven-development principle, the classifier ships
with an eval set (../corpus/failures.jsonl) and a pass threshold. The release
gate is: safety false-positive rate must stay at 0% (no real defect
classified as flaky/environment) and overall accuracy ≥ 90%. The current build
measures 90.7% accuracy / 0% safety-FP (see ../eval/report.md).
In Agent Builder these same cases load as the agent's evaluation set so the agent
cannot be published if it regresses below threshold.