Skip to content

coding-agents.md

Latest commit

 

History

History
84 lines (69 loc) · 4.7 KB

coding-agents.md

File metadata and controls

84 lines (69 loc) · 4.7 KB
Raw

Building FlakeWarden with a coding agent (UiPath for Coding Agents)

This solution was built using Claude Code driving the UiPath uip CLI. This page documents how, so the judges can see the coding-agent workflow (the AgentHack bonus, scored under Platform Usage).

Note on exact commands: uip subcommand names below illustrate the documented UiPath for Coding Agents capability. Confirm the precise flags against your tenant's installed CLI version (uip --help); the build loop is the same.

What the coding agent did

  1. Scaffolded the deterministic core — the feature extractor (features.py), the weighted scorer (scorer.py), and the data schema, from a natural-language spec of the flake signals.
  2. Wrote the grounded classifier with two interchangeable backends (offline rule-based + live Claude), sharing one prompt with the Agent Builder agent.
  3. Generated the evaluation corpus and harness, then iterated on the orchestration when the first eval run exposed a design flaw (environment failures were being auto-classified as defects). The fix — only auto-resolving a defect on a positive selector-change fingerprint, and routing the rest to the grounded classifier — took accuracy from 70% to ~95%. A subsequent self-review caught that the corpus was encoding the exact signals the model read back, so the corpus was decoupled (noisy selector flag, paraphrased error vocabulary, conflict cases) and re-measured at an honest 90.7% with a 0% safety false-positive rate earned by a mechanism rather than the data.
  4. Authored the Maestro process and Agent Builder definitions, then packaged and deployed via the uip CLI.

The build loop (actually executed)

Claude Code drove the real uip CLI against the live Labs tenant. These commands were run, not illustrated:

# set up UiPath for Coding Agents
npm install -g @uipath/cli            # uip 1.196.0
uip login                             # -> jonathansolvesproblems / DefaultTenant
uip skills list                       # 21 UiPath skills for coding agents
uip tools install @uipath/maestro-tool @uipath/agent-tool @uipath/solution-tool \
                  @uipath/orchestrator-tool @uipath/tasks-tool @uipath/test-manager-tool

# deploy the published Agent Builder agent as a runnable Orchestrator process
uip solution packages list            # -> Solution 1, PackageVersionKey 17d2de71...
uip agent deploy 17d2de71-24a1-4d2e-824f-dbcd313ac365 --name FlakeWardenClassifier
  -> Installed -> folder: solution_folder ; process: Solution.1.agent.Agent (ProcessType: Agent)

# author the Maestro BPMN orchestration, registry-driven
uip maestro bpmn registry pull
uip maestro bpmn registry get Orchestrator.StartAgentJob   # agent-call node template
uip maestro bpmn registry get Actions.HITL                 # human-task node template
uip maestro bpmn init flakewarden-maestro --process-id FlakeWardenTriage
  ... Claude Code authors flakewarden-maestro/flakewarden-maestro.bpmn from the
      templates + the structural-BPMN contract (Start -> agent -> verdict script
      -> exclusive gateway -> flaky/real_defect/environment branches + diagram) ...
uip maestro bpmn validate flakewarden-maestro/flakewarden-maestro.bpmn
  -> Status: Valid   (1 process, 1 start event, 12 UiPath extensions, 0 warnings)

The point judges should take away: the coding agent does not just generate code — it drove the platform. It deployed a published agent to Orchestrator and authored a registry-valid Maestro BPMN orchestration end to end through the uip CLI, the exact "build, deploy, operate via natural language" loop UiPath for Coding Agents is for.

What else the coding agent verifiably did in this repo (visible in git history): built the deterministic scorer, the grounded classifier, the eval harness and corpus, then iterated when the eval surfaced a design flaw (environment failures auto-classified as defects, 70%→90.7%) and again when a self-review caught a circular-corpus problem.

Honest remaining step: running the BPMN fully unattended needs the Orchestrator job-argument envelope for the agent process resolved and a serverless robot assigned to solution_folder (the agent runs correctly today in Agent Builder; this is deployment plumbing). The in-Maestro human gate needs an Action Center action app.

Coded + low-code, together

  • Coded: the deterministic scorer and orchestration logic (exact, testable, version-controlled) — authored by the coding agent.
  • Low-code: the Triage Classifier and Repair Agent — Agent Builder agents with grounded sources, output schemas, guardrails, and an eval gate.

This blend is exactly what the hackathon's "combine coding agents with low-code components" bonus rewards.