Hi HN!

We just open-sourced [AdaptiQ Core](https://github.com/adaptiq-ai/adaptiq), a CLI tool that uses reinforcement learning (Q-learning) to optimize your LLM agents and reduce token usage, retries, and failed outputs.

It observes your agent runs, builds a local Q-table, and learns how to improve prompts/configs — all offline.

---

*What it does:*

• Prompt & agent optimizer (crewAI-compatible) • RL loop (offline Q-learning) • Pre-run cost prediction • FinOps reporting (token, $ and CO₂) • Markdown reports + GitHub badge • Works via CLI (`wizard`, `validate`, `report`)

---

*FinOps in Action* Saved Tokens: –37% Retries reduced: –60% Compile-pass +15 pts

GitHub: https://github.com/adaptiq-ai/adaptiq

Thanks for checking out AdaptiQ!

Here’s what’s next in the coming days:

Today: AdaptiQ Core (what you see now) – CLI-based prompt/agent optimizer – Reinforcement Learning loop (offline Q-table) – Token/cost/CO₂ tracking per run – Markdown + badge reporting – Works with CrewAI + OpenAI (GPT-4, GPT-3.5)

Next Week: AdaptiQ ACE – HTTP Proxy Edition – Drop-in FinOps proxy for Claude, Gemini, GPT – Rewrites prompts on-the-fly – Tracks latency, compile pass, retries – GitHub Action: block PR if cost/test fails – RL reward = quality − β·tokens − γ·latency

What we’re solving: > Agents fail silently, burn credits, drift from style guides. > AdaptiQ gives your LLM prompts a feedback loop.

We’re building this in the open: roadmap, CLI, and future trace-spec are all public.

Questions? Feedback? Want support for LangChain / Autogen / Mistral? Let us know below – we’d love to expand!

Also: if you drop your prompt logs (token usage + outcome), we can pre-train the Q-table for your setup.

Cheers – Wassim / AdaptiQ team