Kestrel — Token-Efficient General-Purpose Agent

Track 1 entry for the AMD Developer Hackathon: ACT II.

A containerized agent that answers natural-language tasks across eight capability categories — factual knowledge, math word problems, sentiment classification, summarization, entity extraction, code debugging, logical puzzles, and code generation — engineered to pass the accuracy gate while spending as few Fireworks AI tokens as possible.

Kestrel treats tokens as a budget: it determines how much work each task actually needs and spends exactly that. Every configuration was benchmarked and graded offline before submission, and the container is hardened against API failures, malformed responses, and oversized workloads — it always emits valid, complete results.

Run

The published image implements the Track 1 harness contract (reads /input/tasks.json, writes /output/results.json, exits 0):

docker run --rm \
  -v "$PWD/input:/input:ro" -v "$PWD/output:/output" \
  -e FIREWORKS_API_KEY=... \
  -e FIREWORKS_BASE_URL=... \
  -e ALLOWED_MODELS=... \
  ghcr.io/py-tr/kestrel:a2

Image: ghcr.io/py-tr/kestrel (public, linux/amd64, built and tested by CI on every change).

Source

The full agent source will be published in this repository at submission finalization, ahead of judging. It is withheld during the live competition window for competitive reasons — the runnable artifact above is the complete, scoreable deliverable.

License

MIT — see LICENSE.