agentic-in/inferoa

Inference-native Tokenmaxxing Agent Harness for Loop Engineering

What it solves

Inferoa addresses the inefficiencies that occur during long-horizon agentic loops. As AI agents perform recursive tasks, they often suffer from "prompt drift," collapsing cache reuse, and bloated context windows filled with stale evidence, which increases costs and slows down performance.

How it works

It provides an inference-native runtime that treats the agent loop as an inference workload. It optimizes the process through "tokenmaxxing," which involves preserving cacheable prompt prefixes, bounding mutable context, and using intelligent model routing (via vLLM) to select the most cost-effective model path based on the current task's needs.

Who it’s for

Developers and engineers building complex, recursive AI agents that require durable loops for inspection, editing, testing, and verification without wasting tokens or losing context continuity.

Highlights

  • Loop Engineering: A /loop command that maintains an objective across multiple tasks and attempts until the work is proven.
  • Tokenmaxxing: Tools to monitor and optimize token pressure, prefix-cache reuse, and context savings.
  • Inference-native Routing: Integration with vLLM to route requests based on cost, safety, privacy, and capability.
  • Context Control: Uses compression, summaries, and graph-shaped repository context to keep the window clean of stale state.