Micro-Agentic Signal Loop System

A production-grade, multi-agent orchestration system built on a strict operational constraint: Test reality → Extract signal → Classify → Act.

This is not a chatbot. It is a decision-enforcement engine. It prevents building without signal, ensuring that actions (Scale, Reframe, Kill) are driven by verifiable reality rather than intuition.

Core Operating Philosophy

The system enforces the "Signal Loop" via a sequential, multi-agent pipeline:

• Test Reality (ResearchAgent): Analyzes inputs/files and extracts pure, factual observations. No assumptions. No recommendations.
• Extract Signal (AnalystAgent): Evaluates the research and classifies signal strength (strong, weak, none) using a strict schema tool.
• Determine Action (ActionAgent): Maps the classified signal to an immutable action constraint:
  • Strong → Scale or refine. Build only to extend.
  • Weak → Reframe or reposition. Do not build.
  • None → Change direction. Kill it, test anew.
• Persist (MemoryAgent): Logs the completed loop to context memory so the system compounds intelligence over time.

Production-Grade Features

• Model Agnostic & Future-Proof: Automatically detects and adapts parameter schemas (max_tokens vs max_completion_tokens) for next-gen models (GPT-4o, GPT-4.5, GPT-5.5, o1, o3).
• Multi-Modal File Handling: Natively reads and extracts Text, Word (.docx), PDF (via pypdf), and Images (automatically routed to Vision API).
• Resilient Execution: Wraps LLM calls in exponential backoff retries (Tenacity) for transient network and rate-limit errors.
• Strict Tooling: Agents interact with the environment via pure Python functions mapped to strict JSON schemas.
• In-Memory Context Window: Agents learn from prior tasks via a bounded FIFO memory queue injected into their system prompts.
• Beautiful CLI: Interactive REPL with Markdown rendering, file-upload wizards, and real-time step visualization using Rich.

Project Structure

project/
├── .env
├── requirements.txt
├── src/
    ├── config.py
    ├── memory.py
    ├── tools.py
    ├── agents.py
    ├── orchestrator.py
    └── main.py

Usage

# Install
pip install -r requirements.txt

# Configure
echo "OPENAI_API_KEY=your_key" > .env
echo "OPENAI_MODEL=gpt-4o" >> .env

# Run interactive
python src/main.py

# Run single task
python src/main.py --task "Evaluate demand signal for AI enrichment pipelines targeting family offices"

What Each Component Demonstrates

Component	Production Principle
config.py + Pydantic	Validated config, zero hardcoding
memory.py	Bounded stateful memory — swappable to DB
tools.py	Typed tool registry, clean executor router
agents.py BaseAgent	Retry logic, tool-call loop, shared client
Specialized agents	SRP — one agent, one job, one system prompt
orchestrator.py	Enforces loop order, injects memory context
main.py	CLI + interactive + graceful error handling

---

Extension Paths (Signal-Justified Only)

Strong signal received → build these extensions:
├── Swap memory.py → PostgreSQL (asyncpg)
├── Add web_search tool (Tavily/Serper API)
├── Add send_email tool (ConvertKit API)
├── Add FastAPI wrapper → HTTP endpoint
└── Add LangSmith tracing → observability

---

Outputs

Task: Evaluate demand signal for AI enrichment pipelines targeting family offices