penny-agent

Penny Agent is a general-purpose AI agent that answers natural-language tasks across all eight capability categories — factual knowledge, mathematical reasoning, sentiment classification, summarization, named-entity recognition, code debugging, logical reasoning, and code generation — while minimizing external API cost. Its core is a free, local router. Each incoming task is classified in-process using lightweight keyword heuristics (zero tokens), then dispatched along one of two paths. "Easy" categories that a compact model handles reliably (factual, sentiment, summarization, NER) are answered entirely by a bundled local model — Qwen2.5-3B-Instruct (Q4 GGUF via llama.cpp) — costing zero Fireworks tokens. Correctness-critical categories (math, logic, debugging, code generation) are routed to the appropriate Fireworks model with tightly capped output budgets. A two-way fallback guarantees no task is ever dropped: if one path fails, the other takes over. The design targets the scoring rubric directly — clear the accuracy gate first, then rank by fewest tokens. The model runs comfortably within the 4 GB / 2 vCPU grading VM, completing all tasks in under a minute. Every routing decision, model mapping, and token cap is environment-configurable, so more categories can be shifted local to drive token cost toward zero without code changes.