agent-observer

AgentObserver

Real-time observability, auto-debugging, and evaluation toolkit for AI agents.

---

Inspired by the fact that 40% of enterprise AI agent projects get cancelled — not because of model quality, but due to unobservable production failures. (Gartner Hype Cycle for Agentic AI, 2026)

---

The Problem

AI agents fail silently in production. A loop here, a timeout there, a hallucination nobody caught — and suddenly your agent has rebooked 1,247 passengers onto wrong flights (Air Canada, Jan 2026).

Companies deploying agents have no easy way to:

• Know when an agent is stuck or failing
• Understand why it failed
• Verify if the output is grounded or hallucinated
• Compare which agent performs better
• Evaluate if their RAG pipeline is faithful

AgentObserver solves all of this.

---

Installation

pip install agent-observer

With LangChain support:

pip install agent-observer[langchain]

---

Features

---

Quick Start

1. Basic Agent Monitoring

from agent_observer import AgentObserver

observer = AgentObserver(
    agent_name="MyAgent",
    timeout_threshold=10.0,
    max_retries=3,
    retry_delay=1.0,
)

# Wrap any function call
result = observer.monitor_step("web_search", search_fn, "AI trends 2026")
result = observer.monitor_step("summarize", summarize_fn, result)

# Save session log
summary = observer.save_log("logs/session.json")

Output when failure detected:

=======================================================
  AGENT OBSERVER ALERT — MyAgent
=======================================================
  Step     : #2 — web_search
  Status   : LOOP_DETECTED
  Reason   : Agent repeated 'web_search' 3x with identical input
  Retries  : 0
  Duration : 234ms
=======================================================

---

2. Auto Retry with Exponential Backoff

observer = AgentObserver(
    agent_name="MyAgent",
    max_retries=3,
    retry_delay=1.0,  # doubles each retry: 1s → 2s → 4s
)

# Automatically retries on failure
result = observer.monitor_step("api_call", flaky_api, "query")
# ⚠️ Step #1 failed — retrying in 1.0s... (attempt 1/3)
# ⚠️ Step #1 failed — retrying in 2.0s... (attempt 2/3)
# ✅ Step #1 recovered after 2 retries!

---

3. Hallucination Scorer

from agent_observer import HallucinationScorer

scorer = HallucinationScorer()
report = scorer.score_and_print(llm_output, label="My Output")

# Score      : 92/100  ✅ LOW RISK
# Score      : 30/100  🔴 HIGH RISK
# Score      : 12/100  🚨 CRITICAL RISK

Detects:

• Vague language — "maybe", "possibly", "I think"
• Contradictions — "I know" + "I don't know" in same output
• Overconfidence — "definitely", "certainly" without evidence
• Missing specifics — no dates, numbers, or facts
• Filler density — verbose output with low information

---

4. RAG Monitor

from agent_observer import RAGMonitor

monitor = RAGMonitor()
report = monitor.evaluate(
    question="What is agentic AI?",
    retrieved_docs=["doc1 text...", "doc2 text..."],
    generated_answer="Agentic AI refers to..."
)
report.summary()

Evaluates 4 dimensions:

• Retrieval Quality — are docs relevant to the question?
• Faithfulness — is the answer grounded in docs or hallucinated?
• Answer Relevance — does the answer address the question?
• Hallucination Score — overall output quality

============================================================
📚  RAG MONITOR REPORT
============================================================
  Retrieval Quality : 75/100
  Faithfulness      : 72/100
  Answer Relevance  : 100/100
  Hallucination     : 83/100
────────────────────────────────────────────────────────────
  Overall RAG Score : 82/100  ✅ LOW RISK
============================================================

---

5. Agent Comparison

from agent_observer import AgentComparator

comparator = AgentComparator("GPT-Agent", "Claude-Agent")
comparator.add_task("web_search", gpt_search, claude_search, "AI trends")
comparator.add_task("summarize", gpt_summarize, claude_summarize, "text here")

report = comparator.run_and_compare()

Output:

============================================================
  COMPARISON REPORT
============================================================
  METRIC              GPT-Agent        Claude-Agent
  Failure Rate     👑  0.0%            20.0%
  Avg Duration     👑  320ms           720ms
  Total Retries    👑  0               5
  Weighted Score   👑  100/100         20/100
────────────────────────────────────────────────────────────
    OVERALL WINNER: GPT-AGENT
============================================================

---

6. LangChain Plugin

from agent_observer import AgentObserverCallback
from langchain.agents import initialize_agent

# Just add one line to any existing LangChain agent
callback = AgentObserverCallback(
    agent_name="MyLangChainAgent",
    timeout_threshold=10.0,
    hallucination_threshold=50,
)

agent = initialize_agent(
    tools=tools,
    llm=llm,
    callbacks=[callback]  # ← that's it!
)

agent.run("Your task here")

Automatically monitors:

• Every tool call — failures, timeouts, loops
• Every LLM generation — hallucination scored in real-time
• Full session log saved on completion

---

7. AI Postmortem Report

from agent_observer import print_postmortem

summary = observer.save_log("logs/session.json")
print_postmortem(summary, api_key=GROQ_API_KEY)

Output:

📋 POSTMORTEM REPORT
────────────────────────────────────────
1. Summary
   Agent got stuck in an infinite search loop on Step 4.

2. Root Cause
   Search query was not updated between retries — agent had
   no logic to modify query on repeated failure.

3. Impact
   3 redundant API calls, 700ms added latency, no useful output.

4. Fix Recommendation
   Add query variation logic — if same query fails twice,
   rephrase before retrying. Add max_retries=2 guard.
────────────────────────────────────────

---

8. Dashboard

python examples/demo.py        # generate a session first
streamlit run dashboard/app.py # launch dashboard

Visual analytics including:

• Step log with color-coded pass/fail
• Failure breakdown by type
• Step duration chart
• One-click AI postmortem generation

---

Running the Examples

# Core failure detection + retry
python examples/demo.py

# Agent comparison
python examples/compare_demo.py

# Hallucination scoring
python examples/hallucination_demo.py

# LangChain plugin
python examples/langchain_demo.py

# RAG monitoring
python examples/rag_demo.py

# Dashboard
python -m streamlit run dashboard/app.py

---

Failure Types Detected

---

Project Structure

agent-observer/
├── core/
│   ├── observer.py              # Failure detection + retry engine
│   ├── postmortem.py            # AI postmortem generator (Groq)
│   ├── comparator.py            # Agent comparison
│   ├── hallucination_scorer.py  # Hallucination scoring
│   ├── langchain_plugin.py      # LangChain callback plugin
│   └── rag_monitor.py           # RAG pipeline evaluation
├── dashboard/
│   └── app.py                   # Streamlit dashboard
├── examples/
│   ├── demo.py
│   ├── compare_demo.py
│   ├── hallucination_demo.py
│   ├── langchain_demo.py
│   └── rag_demo.py
├── logs/                        # Auto-generated session logs
└── README.md

---

Environment Variables

# For AI postmortem reports (free)
export GROQ_API_KEY=your-groq-key-here

# Windows
set GROQ_API_KEY=your-groq-key-here

Get a free Groq API key at console.groq.com

---

Research Background

This project addresses open research challenges in agentic AI reliability:

• "Agentic Uncertainty Quantification" — arXiv:2601.15703 (2026)
• "Why AI Agents Fail in Production" — Gartner Hype Cycle for Agentic AI (2026)
• "40% of enterprise agentic AI projects expected to fail by 2027" — Gartner (2026)

---

Roadmap

• Core failure detection engine
• Auto retry with exponential backoff
• AI-powered postmortem reports
• Hallucination scorer
• RAG pipeline monitor
• Agent comparison
• LangChain plugin
• Streamlit dashboard
• PyPI package

---

License

MIT License — see LICENSE

---

Author

Built by Mannya

⭐ Star this repo if you find it useful!