TEA11

This is my Track 1 entry. It's a containerized agent that answers a batch of natural-language tasks across all eight categories - factual questions, math, sentiment, summarization, named-entity recognition, code debugging, logic, and code generation - using as few Fireworks tokens as possible. I built the whole thing around the scoring rule: passing submissions are ranked by total tokens, so the goal is correct answers with minimal spend. Instead of heavy prompting or multi-step chains, each task gets a single tightly-scoped model call, and I cut anything that doesn't help produce a correct answer. The model and the API endpoint are both read from the judge's environment at run time, so nothing is hardcoded and it works with whatever allowed models are provided. It runs the batch in parallel to stay well inside the time limit, with small safeguards so one failed or empty response can't sink the whole run. The image is a slim linux/amd64 build with no model weights bundled, so it stays far under the size cap and starts fast.