CarScout

• 

  eventbridge -schedule

• 

  OTEL Session

• 

  lamda function

• 

  observability

• 

  agentcore -runtime

• 

  architecture

• 

  timelines

• 

  request-response

Inspiration

Shopping for a used car is a part-time job on the listings' schedule, not yours: refresh, open twenty tabs, price each against its year and mileage, dodge the salvage titles and wire-transfer scams, message fast before the good ones go.

What it does

Runs on a cron, scores every listing against a fair-price estimate, your hard filters and a red-flag check, and returns a short shortlist with pursue / consider verdicts plus an explicit list of the scams and bad-title cars to avoid. Drafts seller outreach on request. Remembers what you've dismissed.

How we built it

• Strands Agents SDK — one Agent, six @tools, a "protect the buyer's time" system prompt.
• Amazon Bedrock — Claude Haiku 4.5.
• Amazon Bedrock AgentCore — BedrockAgentCoreApp entrypoint, EventBridge cron.
• Deterministic core — fair-price depreciation model, criteria matching and scam heuristics in carscout/core.py, 12 tests, no model calls.

Challenges we ran into

Scam detection that is strict without being paranoid. A CRITICAL flag (bad title or a price far below market) hard-caps the score and forces skip — the LLM can't be talked out of it by a persuasive description.

Accomplishments that we're proud of

The whole valuation and risk layer runs offline (make demo). The agent is a thin triage layer that's easy to trust because it can't do the math wrong.

What we learned

"Only show me what matters" is a prompt problem and a data problem — you need a worth_your_time signal from tested code for the model to lean on.

What's next for CarScout

Live marketplace scraping (Playwright, logged-in session), a real valuation API, VIN history lookups, AgentCore Memory, push delivery.

Built With

• bedrockagentcore
• python
• strandsagents