AI Research Agent

An autonomous research agent powered by Amazon Bedrock that generates intelligent, structured insights on any topic.

Comment

Inspiration

Finding reliable, structured insight fast is hard. I wanted an autonomous “junior researcher” that I could ping with a topic and instantly get a clean brief: summary, key findings, risks, and sources—without standing up servers. AWS Bedrock + Lambda felt perfect to show how far agentic workflows have come.

What it does

AI Research Agent is a serverless agent that:

  • Accepts a topic via a public API (API Gateway).
  • Enriches context with quick web snippets.
  • Uses Amazon Bedrock (Claude 3.5 Sonnet) to reason and write a Markdown research brief.
  • Stores results in S3 (reports/*.md) and returns a 1-hour presigned URL.
  • Every run creates a new, timestamped file—nothing is overwritten.

Typical output sections

  • Summary (3–5 sentences)
  • Key Findings (bullets)
  • Opportunities & Risks
  • Implementation Notes
  • Sources

How we built it

  • API Gateway exposes a simple POST endpoint ({"topic":"..."}).
  • AWS Lambda (Python) is the agent brain: parses the request, optionally calls Serper, prompts Claude, saves to S3, returns JSON + presigned URL.
  • Amazon Bedrock hosts Claude 3.5 Sonnet and handles reasoning/writing.
  • Amazon S3 stores Markdown artifacts in reports/….
  • CloudWatch collects logs for quick debugging.

Key design choices

  • Stateless Lambda → easy to scale and super cheap.
  • Timestamped filenames → durable audit trail.
  • Presigned URLs → frictionless viewing for judges/users.

Challenges we ran into

  • Invocation plumbing: API Gateway wraps the body; added robust parsing so both console tests and API calls work.
  • Model headers/IDs: Bedrock requires correct contentType/accept and model ID suffixes (e.g., -v1:0).
  • Rate limits: occasional ThrottlingException—added retries/manual backoff in testing.
  • Permissions: minimal IAM policy for Bedrock + S3 access without over-permissioning.

Accomplishments that we’re proud of

  • Fully working agentic flow end-to-end on AWS in under a day.
  • Clean, reproducible architecture with almost zero ops overhead.
  • Simple public API that anyone can try with a one-line curl.

What we learned

  • Bedrock’s Anthropic models are excellent for structured outputs with careful prompting.
  • Small details (model IDs, headers, API Gateway body formats) matter.
  • Serverless agents are a sweet spot for productionizing LLM workflows quickly and safely.

What’s next for AI Research Agent

  • Evidence table with citation confidence, and retrieval augmentation from a private corpus.
  • Async jobs + notifications for longer reports.
  • Frontend (tiny React app) and multi-topic batch mode.
  • Agent chaining: plan → research → write → fact-check → publish.

Built With

Share this project:

Updates

Submission history