Hi Product Hunt, I'm Hamed, one of the authors behind AgentCodec. Your phone doesn't send every message the same way. Next to a tower it packs in as much data as it can. In a basement it slows down and adds redundancy so nothing gets lost. Communication engineers have been refining this since Shannon's 1948 paper. Most AI apps don't adapt at all. Every request goes to the same model, called the same way, whether it's trivial or hard. The easy ones are overpaid for, and the hard ones still fail. Our paper, "A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability," was just accepted to NeurIPS 2026. It shows that an LLM behaves like a noisy channel in the mathematical sense, so the same toolkit applies. On the paper's text benchmarks, our router cut cost by about 56% at the same quality as the best fixed technique, or raised quality by about 7% at the same cost. It doesn't just pick a point on the cost-quality frontier; it draws a new one. The paper was the starting point. AgentCodec has improved a lot since, and it now works with vision-language models too. On MMMU, two open-weight VLMs combined through AgentCodec scored clearly higher than either one alone with thinking on. What you get: - 28 reliability techniques: 21 from communication theory (HARQ, turbo, fountain, FEC, diversity combining, adaptive routing) and 7 well-known ones such as self-consistency and best-of-N - Evaluations on a sample of your own data, across the models you choose: hosted or open-weight, text or vision - A router that picks the model and technique for each request, with one knob for cost vs. quality (on the Dedicated plan you can train it on your own runs) - One endpoint, OpenAI-compatible if you want it One small example from our vision tests: an MMMU question asks how many measures are in a short music score that has no bar lines. Both thinking models we tested got it wrong, and one of them spent over 26,000 tokens counting. AgentCodec's HARQ technique, the same retry scheme your phone uses on a weak signal, got it right for the same cost. AgentCodec is available now on the AptAgents platform, with a 7-day free trial on the Standard plan. More on the product page: https://intellerce.com/products/... We'd value your feedback, especially if you run agents in production. Where do your requests fail today, and what would you need to trust a router with them? I'll be here all day to answer questions.

AgentCodec

Combine cheaper models to rival frontier LLMs and VLMs

Combine cheaper models to rival frontier LLMs and VLMs