Strategies
Wiyomi uses published cost-saving methods inside the agent. Results below are from the original papers.
01
AgentDiet removes useless, redundant, and expired context from a coding agent's trajectory while it runs.
Result
39.9–59.7% fewer input tokens and 21.1–35.9% lower cost, with the same agent performance.
arXiv:2509.2358602
Tests repository-level instructions on coding agents working through GitHub pull requests.
Result
28.64% lower median runtime and 16.58% fewer output tokens, with comparable task completion.
arXiv:2601.2040403
Measures where tokens go in agentic coding, and whether models can predict that cost before a run.
Result
Agentic coding can use about 1,000× more tokens than code chat. More tokens do not mean higher accuracy, and frontier models underestimate their own cost.
arXiv:2604.2275004
Gives an agent a running tool-call budget, then adapts planning and verification as that budget is used.
Result
Budget-aware agents scale more favorably than agents that simply receive a larger tool-call budget.
arXiv:2511.1700605
Estimates how many reasoning tokens a query needs, then keeps the response within that budget.
06
Sends each query to a stronger or weaker model, using preference data to make the choice.
07
Combines routing and cascading, so a query can skip models or stop once the answer is good enough.
Result
Outperforms routing alone and cascading alone on the cost–performance tradeoff.
arXiv:2410.1034708
Compresses long prompts before inference while keeping the instruction and task meaning.
09
Compresses prompts without a specific task, and keeps the shorter prompt faithful to the original.
Result
3–6× faster than earlier compressors, and 1.6–2.9× faster end to end at 2–5× compression.
arXiv:2403.1296810
Measures when prompt compression reduces real latency, not only the token count.
Result
Up to 18% faster end to end when prompt length, ratio, and hardware match. Quality stayed statistically unchanged on summarization, code generation, and question answering.
arXiv:2604.0298511
Separates prompt compression from response compression, and scores the bill that is actually paid.
Result
Output compression cuts realized cost by 1.4–2.4× on most API models. Naive input compression can raise cost and reduce accuracy.
arXiv:2606.2408312
Caches the intermediate state of retrieved documents and reuses it across requests.
13
Reuses cached knowledge even when the retrieved text is not a prefix, then recomputes a small set of tokens.
Result
2.2–3.3× faster time to first token and 2.8–5× higher throughput, without compromising generation quality.
arXiv:2405.16444