Routing-Agent-That-Works-I-Guess?

Routing-Agent-That-Works-I-Guess? is a token-efficient hybrid agent built for AMD Hackathon Track 1. It answers eight general-purpose task types—factual knowledge, math, sentiment, summarization, NER, code debugging, logic puzzles, and code generation—while keeping Fireworks API usage as low as possible. The agent follows a deterministic-first strategy. Each task is classified by category, output format, and risk level, including a freshness gate for current or hallucination-sensitive prompts. When pattern matching and local solvers can answer confidently, the agent validates locally and uses zero Fireworks tokens. For remaining low- and medium-risk tasks, it can draft answers locally and send them through a compact Fireworks validate/fix step. High-risk or time-sensitive tasks skip weak local paths and go directly to Fireworks. The container reads /input/tasks.json, processes every task with per-task fallbacks and schema validation, and writes /output/results.json before exiting successfully. Model selection is driven entirely by runtime environment variables, and the image is built for linux/amd64 to meet judging requirements. The result is a practical smart-router pattern: strong accuracy where it matters, with API cost controlled through routing rather than always calling a large model.