agent-reasoning 1.0.7

Transform LLMs into robust problem-solving agents with advanced reasoning strategies

Agent Reasoning: The Thinking Layer

Vision & Purpose

The Reasoning Layer is the cognitive engine of the AI stack. While traditional LLMs excel at token generation, they often struggle with complex planning, logical deduction, and self-correction.

This repository transforms standard Open Source models (like gemma3, llama3) into robust problem solvers by wrapping them in advanced cognitive architectures. It implements findings from key research papers (CoT, ToT, ReAct) to give models "agency" over their thinking process.

"From predicting the next token to predicting the next thought."

---

📦 Installation

From PyPI (Recommended)

pip install agent-reasoning

# With server dependencies (for the reasoning gateway):
pip install "agent-reasoning[server]"

From Source

# Clone the repo
git clone https://github.com/jasperan/agent-reasoning.git
cd agent-reasoning

# Install dependencies
python3 -m venv venv
source venv/bin/activate
pip install -e .

Prerequisite: Ollama must be running locally, or you can connect to a remote Ollama instance.

ollama pull gemma3:270m

Configuring Remote Ollama Endpoint

If you don't have Ollama installed locally, you can connect to a remote Ollama instance. Configuration is stored in config.yaml in the root directory of the repository.

Option 1: Interactive CLI Configuration

agent-reasoning
# Select "Configure Endpoint" from the menu

Option 2: Server CLI Argument

agent-reasoning-server --ollama-host http://192.168.1.100:11434

Option 3: Direct Config File

Copy the example config and edit it:

cp config.yaml.example config.yaml

Or create config.yaml in the project root:

ollama:
  host: http://192.168.1.100:11434

Option 4: Python API

from agent_reasoning import ReasoningInterceptor, set_ollama_host, get_ollama_host

# Check current endpoint
print(get_ollama_host())  # http://localhost:11434

# Set a new endpoint (persists to config file)
set_ollama_host("http://192.168.1.100:11434")

# Or specify directly without saving to config
client = ReasoningInterceptor(host="http://192.168.1.100:11434")

---

📓 Notebooks

Interactive Jupyter notebooks demonstrating agent reasoning capabilities:

Name	Description	Stack	Link
agent_reasoning_demo	Comprehensive demo of all reasoning strategies (CoT, ToT, ReAct, Self-Reflection) with benchmarks and comparisons	Ollama, Gemma3/Llama3, FastAPI

---

🚀 Features

✅ Verified against ArXiv Papers

• Plug & Play: Use via Python Class or as a Network Proxy.
• Model Agnostic: Works with any model served by Ollama.
• Advanced Architectures:
  • 🔗 Chain-of-Thought (CoT) & Self-Consistency: Implements Majority Voting ($k$ samples) with temperature sampling.
  • 🌳 Tree of Thoughts (ToT): BFS strategy with robust heuristic scoring and pruning.
  • 🛠️ ReAct (Reason + Act): Real-time tool usage (Web Search via scraping, Wikipedia API, Calculator) with fallback/mock capabilities. External grounding implemented.
  • 🪞 Self-Reflection: Dynamic multi-turn Refinement Loop (Draft -> Critique -> Improve).
  • 🧩 Decomposition & Least-to-Most: Planning and sub-task execution.

---

💻 Usage

1. Interactive CLI (Recommended)

Access all agents, comparisons, and benchmarks via the rich CLI.

# If installed via pip:
agent-reasoning

# Or from source:
python agent_cli.py

Interactive Experience:

╭────────────────────────────────────────────╮
│ AGENT REASONING CLI                        │
│ Advanced Cognitive Architectures (Gemma 3) │
╰────────────────────────────────────────────╯

? Select an Activity:
  Chat with Standard Agent
  Chain of Thought (CoT)
  Tree of Thoughts (ToT)
  ReAct (Tools + Web)
  ⚔️  ARENA: Run All Compare
  Select AI Model
  Exit

2. Python API (For Developers)

Use the ReasoningInterceptor as a drop-in replacement for your LLM client.

from agent_reasoning import ReasoningInterceptor

client = ReasoningInterceptor()

# Append the strategy to the model name with a '+'
response = client.generate(
    model="gemma3:270m+tot",
    prompt="I have a 3-gallon and 5-gallon jug. How do I measure 4 gallons?"
)
print(response["response"])

Using agents directly:

from agent_reasoning.agents import CoTAgent, ToTAgent, ReActAgent

# Create an agent
agent = CoTAgent(model="gemma3:270m")

# Stream responses
for chunk in agent.stream("Explain quantum entanglement step by step"):
    print(chunk, end="")

3. Reasoning Gateway Server

Run a proxy server that impersonates Ollama. This allows any Ollama-compatible app (LangChain, Web UIs) to gain reasoning capabilities without code changes.

# If installed via pip:
agent-reasoning-server --port 8080

# Or from source:
python server.py

Then configure your app:

• Base URL: http://localhost:8080
• Model: gemma3:270m+cot (or +tot, +react, etc.)

Example:

curl http://localhost:8080/api/generate -d '{
  "model": "gemma3:270m+cot",
  "prompt": "Why is the sky blue?"
}'

---

🧠 Architectures in Detail

Architecture	Description	Best For	Papers
Chain-of-Thought	Step-by-step reasoning prompt injection.	Math, Logic, Explanations	Wei et al. (2022)
Self-Reflection	Draft -> Critique -> Refine loop.	Creative Writing, High Accuracy	Shinn et al. (2023)
ReAct	Interleaves Reasoning and Tool Usage.	Fact-checking, Calculations	Yao et al. (2022)
Tree of Thoughts	Explores multiple reasoning branches (BFS/DFS).	Complex Riddles, Strategy	Yao et al. (2023)
Decomposed	Breaks complex queries into sub-tasks.	Planning, Long-form answers	Khot et al. (2022)
Recursive (RLM)	Uses Python REPL to recursively process prompt variables.	Long-context processing	Author et al. (2025)

---

📚 Appendix A: Extending the System

To add a new reasoning strategy (e.g., "Reviewer-Critic"), simply:

1. Create a class in src/agent_reasoning/agents/ inheriting from BaseAgent.
2. Implement the stream(self, query) method.
3. Register it in AGENT_MAP in src/agent_reasoning/interceptor.py.

from agent_reasoning.agents.base import BaseAgent

class MyNewAgent(BaseAgent):
    def stream(self, query):
        yield "Thinking differently...\n"
        # ... your custom logic ...
        yield "Final Answer"

🔧 Appendix B: Troubleshooting

• Model Not Found: Ensure you have pulled the base model (ollama pull gemma3:270m).
• Timeout / Slow: ToT and Self-Reflection make multiple calls to the LLM. With larger models (Llama3 70b), this can take time.
• Hallucinations: The default demo uses gemma3:270m which is extremely small and prone to logic errors. Switch to gemma2:9b or llama3 for robust results.

---

📊 Benchmark Report (Example Outputs)

Below are real outputs generated by the main.py benchmark using gemma3:270m. Note that while the small model strives to follow the reasoning structures, its logic limitations highlight the importance of using larger models (e.g., llama3 or gemma2:9b) for production.

1. Philosophy (Self-Consistency)

Generates multiple reasoning paths and votes for the best answer.

Query: "What is the meaning of life? Answer with a mix of biological and philosophical perspectives."

[ConsistencyAgent]: Processing query via Self-Consistency (k=3)...
  Sample 1: [Detailed biological perspective on propagation...]
  Sample 2: [Philosophical view on existentialism and purpose...]
  Sample 3: [Synthesis of both views...]
Majority Logic: [Aggregated Best Answer from Votes]

2. Logic (Tree of Thoughts)

Explores multiple branches (BFS) to solve riddles.

Query: "I have a 3-gallon jug and a 5-gallon jug. How can I measure exactly 4 gallons of water?"

[ToTAgent]: Processing query via Tree of Thoughts (BFS)...
Thinking via Tree of Thoughts (Depth=3, Width=2)...

[Step 1/3 - Exploring branches]
  Path Score: 0.0
  Path Score: 1.0

[Step 2/3 - Exploring branches]
  Path Score: 1.0
  Path Score: 1.0
  Path Score: 0.1

[Step 3/3 - Exploring branches]
  Path Score: 1.0 (Found solution state)

[Best Logic Trace selected. Generating Final Answer]
**Final Answer:**
1. Pour water from the 5-gallon jug into the 3-gallon jug.
2. You now have 2 gallons left in the 5-gallon jug.
3. Empty the 3-gallon jug.
4. Pour the 2 gallons from the 5-gallon jug into the 3-gallon jug.
5. Fill the 5-gallon jug again.
6. Pour from the 5-gallon jug into the 3-gallon jug until full (needs 1 gallon).
7. You are left with exactly 4 gallons in the 5-gallon jug.

3. Planning (Decomposed Agent)

Breaks down complex tasks into sub-problems.

Query: "Plan a detailed 3-day itinerary for Tokyo for a history buff who loves samurais and tea."

[DecomposedAgent]: Decomposing the problem...

Sub-tasks Plan:
1.  **Define the Scope:** What historical period and specific area of Tokyo will the itinerary cover?
2.  **Identify Key Historical Sites:** What historical sites will the itinerary focus on?
3.  **Determine Traveler's Interests:** What types of historical sites will the itinerary include?
4.  **Outline the Itinerary:** What activities and attractions will be included in each day?
5.  **Estimate Duration:** How long will the itinerary last?

[DecomposedAgent]: Solving sub-task: 1. Define the Scope...
[DecomposedAgent]: Solving sub-task: 2. Identify Key Historical Sites...
...
Final Answer: [Detailed 3-day plan covering Meiji Shrine, Tea Ceremonies, and Samurai Museum]

4. Tool Use (ReAct Agent)

Interleaves thought, action, and observation to solve problems.

Query: "Who is the current CEO of Google? Calculate the square root of 144."

[ReActAgent]: Processing query with ReAct...

--- Step 1 ---
Agent: Action: web_search[current CEO of Google]
Observation: Sundar Pichai is the current CEO of Google.
Final Answer: Sundar Pichai

Running web_search...
Observation: [1] Sundar Pichai - Wikipedia: ... He is the chief executive officer (CEO) of Alphabet Inc. and its subsidiary Google.

---

📄 License

MIT License - see LICENSE for details.