@reactive-agents/reactive-intelligence

Reactive intelligence layer for the Reactive Agents framework.

Real-time entropy sensing, adaptive control, and local learning that monitors agent reasoning quality and intervenes automatically — triggering early stops, context compression, or strategy switches when the agent is struggling.

Installation

bun add @reactive-agents/reactive-intelligence

Or install everything at once:

bun add reactive-agents

Usage

import { ReactiveAgents } from "reactive-agents";

const agent = await ReactiveAgents.create()
  .withName("smart-agent")
  .withProvider("anthropic")
  .withReasoning()
  .withReactiveIntelligence()
  .build();

const result = await agent.run("Analyze market trends for Q1 2026");

Direct Service Usage

import {
  EntropySensorServiceLive,
  ReactiveControllerServiceLive,
  LearningEngineServiceLive,
  createReactiveIntelligenceLayer,
} from "@reactive-agents/reactive-intelligence";

Architecture

Entropy Sensor

Scores reasoning quality each iteration using 5 entropy sources:

| Source | What It Measures | | --- | --- | | Token entropy | LLM confidence from logprobs | | Structural entropy | Format compliance, hedge words, thought density | | Semantic entropy | Task alignment, novelty, repetition | | Behavioral entropy | Tool reuse, action diversity, loop detection | | Context pressure | Context window utilization and overflow risk |

Reactive Controller

Consumes entropy scores and makes real-time decisions:

• Early stop — halt when entropy is consistently low (agent is confident)
• Context compression — trigger compaction when context pressure is high
• Strategy switch — switch reasoning strategy when the current one is looping

Learning Engine

Improves agent performance over time without additional prompting:

• Conformal calibration — adjusts entropy thresholds to the specific model
• Thompson Sampling bandit — learns which strategies work best per task category
• Skill synthesis — extracts reusable skill fragments from successful runs

Telemetry Client

Opt-in telemetry that reports anonymized run metrics (HMAC-signed, fire-and-forget) to improve the framework.

Key Features

• 5-source entropy scoring — composite signal with adaptive weights per model
• Trajectory analysis — classifies entropy trends as converging, flat, diverging, oscillating, or v-recovery
• Model registry — pre-calibrated baselines for common models (Claude, GPT-4, Gemini, Ollama)
• EventBus-driven — entropy scores publish as events, no polling required
• Zero config — .withReactiveIntelligence() enables everything with sensible defaults

Documentation

Full documentation at docs.reactiveagents.dev

License

MIT