Skip to content

Challenge 008: Build a Local-Only Agent

Level: L200 Type: Challenge Time: ~75 min 💰 Cost: Free (local)

Scenario

OutdoorGear wants a tiny support agent that works without Azure, GitHub Models, hosted LLMs, or network calls. It should classify a request, use local tools over local data, and return a grounded answer with a trace.

Your job is to finish the local-only agent and prove that it can handle product search, policy lookup, and bundle recommendation requests without cloud configuration.

Objective

Fix starter_local_agent.py so the local agent routes requests correctly, uses only local data, returns useful answers, reports no cloud usage, and generates a validation code.

Your final local agent should:

  • Require no cloud configuration
  • Classify three supported intents
  • Search products locally
  • Look up policy text locally
  • Build a camping bundle within budget
  • Return answer, trace, intent, and uses_cloud: False

Starter Files

Save these files in one folder named challenge-008/:

File Purpose Download
local_data.json Local product and policy data Download
eval_requests.json Evaluation requests Download
starter_local_agent.py Broken local-only agent Download
test_local_agent.py Acceptance tests Download
validate_local_agent.py Generates the final completion code Download

Challenge Brief

You receive local product/policy data, evaluation requests, and a broken local agent. There is no walkthrough: decide how to route intents, call local tools, compose answers, and prove that no cloud dependency is required.

Constraints

  • Use only the Python standard library in starter_local_agent.py.
  • Do not call any network, hosted model, or cloud API.
  • Do not hardcode answers by request ID.
  • Preserve the public function names used by the tests.
  • Keep traces useful enough to show which local tool ran.

Acceptance Criteria

Your solution is complete when:

  • python -m pytest test_local_agent.py passes
  • Product search returns RainShell Pro Jacket
  • Policy lookup returns the 60-day return policy
  • Bundle recommendation includes the expected three camping items under budget
  • The trace records intent and tool execution
  • uses_cloud is False

Validation

When your implementation is ready, run:

python -m pytest test_local_agent.py
python validate_local_agent.py

Enter the completion code printed by validate_local_agent.py:

Hints

Hint 1 — Local-only is a requirement

A local agent can still be useful if intent routing and tools are deterministic.

Hint 2 — Keep routing narrow

The fixture requests are intentionally scoped to product search, policy lookup, and bundle recommendation.

Hint 3 — Trace the tool, not just the answer

A trace helps prove that the agent made a decision and used a local tool.

Rubric

Area Points What good looks like
Local-only behavior 25 No cloud or network dependency
Intent routing 25 Three request types classified correctly
Tool outputs 25 Product, policy, and bundle answers are correct
Traceability 15 Trace shows intent and selected local tool
Simplicity 10 Deterministic local implementation