Cognitive Kernel-Pro: A Framework for Deep Research Agents and Agent Foundation Models Training

📝 Paper Summary

Agentic AI frameworks Open-source agents Agent foundation models

Cognitive Kernel-Pro is an open-source, multi-module agent framework that minimizes reliance on paid tools and introduces a comprehensive training recipe for deep research agents.

Core Problem

Current advanced AI agent systems heavily rely on closed-source models or paid proprietary tools (like Jina Reader or FireCrawl) to achieve competitive performance.

Why it matters:

Dependency on paid tools creates significant barriers to accessibility and reproducibility for the broader research community
Existing open-source frameworks often lack multimodal or general agentic abilities without these external dependencies

Concrete Example: When performing web navigation or file processing, existing open-source agents often rely on proprietary parsing tools like Chunkr or FireCrawl, whereas Cognitive Kernel-Pro handles these autonomously using its native Web and File sub-agents without external paid dependencies.

Key Novelty

Fully Open-Source Hierarchical Agent Framework (Cognitive Kernel-Pro)

Implements a two-tier hierarchical framework where a main agent orchestrates specialized sub-agents (Web, File) using Python code as the universal action space
Minimizes paid dependencies to only the Google Search API, avoiding proprietary parsing tools to maximize accessibility and reproducibility
Introduces a data synthesis pipeline using intermediate process hints and Persona Hub to generate verifiable trajectories for training agent foundation models

Architecture

The two-tier multi-module architecture of the Cognitive Kernel-Pro framework, detailing the interaction between the main agent and specialized sub-agents

Evaluation Highlights

Surpasses the open-source Smolagents framework by 5% in Pass@1 and 7% in Pass@3 on the GAIA (General AI Assistants) benchmark using Claude-3.7
The fine-tuned CK-Pro-8B model achieves a 38.18% Pass@3 score on GAIA, establishing a new standard for accessible 8B-parameter agents
Yields the best Pass@1 and Pass@3 performance across all levels of the text-only GAIA subset compared to prior open models like WebDancer and WebSailor

Breakthrough Assessment

8/10

Provides a robust, fully open-source alternative to proprietary agent frameworks, achieving state-of-the-art results among free agents while contributing a valuable data synthesis methodology.

⚙️ Technical Details

Problem Definition

Setting: General-purpose task solving and deep research via multi-step reasoning, web navigation, and file processing

Inputs: A natural language task string (e.g., deep research query)

Outputs: A final response string, with intermediate reasoning executed as Python code

Pipeline Flow

Main Agent (Task Decomposition & Delegation)
Web Agent / File Agent (Specialized Sub-task Execution)
Main Agent (Information Aggregation & Final Response)

System Modules

Main Agent

Decomposes complex tasks, delegates sub-tasks to sub-agents, aggregates information, and generates Python code for tool calling and reasoning

Model or implementation: Shared Agent Foundation Model (e.g., CK-Pro-8B, Claude-3.7, or GPT-4.1)

Web Agent (Environment Interaction)

Navigates live web pages using Playwright to collect time-sensitive information

Model or implementation: Shared Agent Foundation Model

File Agent (Environment Interaction)

Processes various file types (PDF, spreadsheet, images) by splitting them into readable pages

Model or implementation: Shared Agent Foundation Model

Novel Architectural Elements

Two-tier hierarchical architecture decoupling task decomposition (Main Agent) from resource-specific execution (Sub-Agents)
Universal Python code action space for all agents, making every sub-agent inherently a code-generation agent

Modeling

Base Model: Qwen-3-8B (for the open-source CK-Pro-8B agent foundation model)

Training Method: Supervised Fine-Tuning (SFT) on synthesized agent trajectories

Training Data:

Multi-hop Web Search Data: Agent traverses seed URLs to compose complex questions
Agent Exploration-based Data: Uses intermediate results as textual hints (wrapped in <secret> tags) to improve trajectory sampling success, removing hints before training
Persona Hub-based Data Augmentation: Synthesizes 1k deep research queries using diverse personas
Reasoning Data: Converted from existing datasets (NumiaMath, LogiCoT, TACO, RiddleSense)

Comparison to Prior Work

vs. Smolagents: Cognitive Kernel-Pro achieves higher Pass@1 and Pass@3 on GAIA under identical experimental conditions
vs. OWL: Achieves rival performance without relying on proprietary file/web processing tools
vs. WebDancer & WebSailor: CK-Pro-8B surpasses their performance on the text-only GAIA subset
+ 1 more
vs. Search-o1: CK-Pro provides a multi-module architecture handling local files and multimodal inputs natively, whereas Search-o1 focuses primarily on text-based web search [not cited in paper]

Limitations

Still relies on one paid tool (Google Search API), preventing the framework from being 100% free
Performance of the open-source CK-Pro-8B model still lags significantly behind state-of-the-art proprietary models like Claude-3.7 (a 30% gap on GAIA)
Persona Hub synthetic queries lack manual validation and response annotation, relying solely on cross-validation of trajectory outcomes
No statistical significance tests reported for benchmark improvements

Reproducibility

Code: https://github.com/Tencent/CognitiveKernel-Pro

Code is publicly available at GitHub. Relies on Google Search API (paid) but avoids other proprietary tools like Chunkr or FireCrawl. Uses GPT-4.1 for trajectory generation. Training data construction methodology is detailed, but specific hyperparameter values for fine-tuning the 8B model are missing from the text.

📊 Experiments & Results

Evaluation Setup

General intelligence and multi-step reasoning evaluation across diverse tasks (web navigation, question answering, file manipulation)

Benchmarks:

GAIA (General AI agent benchmark (web, file, multimodal reasoning))

Metrics:

Pass@1
Pass@3
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Cognitive Kernel-Pro utilizing Claude-3.7 outperforms the Smolagents framework by 5% in Pass@1 and 7% in Pass@3 on the complete GAIA dataset under identical experimental conditions
The framework achieves performance rivaling OWL, despite OWL's heavy reliance on proprietary parsing tools like Chunkr and FireCrawl
The fine-tuned CK-Pro-8B model achieves a Pass@3 of 38.18% on GAIA, and yields the best performance across all levels of the text-only GAIA subset compared to baselines like WebDancer and WebSailor
Ablation studies show that adding reflection ability during test-time significantly improves performance for the trained CK-Pro-8B model, indicating a strong future direction for agent foundation model training
Replacing the multimodal backbone Qwen-2.5-VL-72B with GPT-4.1 yields only marginal improvements, suggesting the open-weight Vision-Language Model is highly competitive for agentic tasks

📚 Prerequisite Knowledge

Prerequisites

Large Language Models (LLMs) and Vision-Language Models (VLMs)
Agentic AI and tool calling architectures
Supervised Fine-Tuning (SFT) for language models

Key Terms

LLMs: Large Language Models—AI models trained on vast amounts of text to understand and generate human language

VLMs: Vision-Language Models—AI models capable of processing both text and image inputs simultaneously

Deep Research Agents: AI agents capable of complex reasoning, web navigation, file processing, and autonomous knowledge discovery

Agent Foundation Models: Base language or multimodal models specifically pre-trained or fine-tuned to act as the core decision-making engine for AI agents

Persona Hub: A strategy to synthesize large-scale diverse queries by using various personas as triggers for an LLM to generate synthetic tasks

Rejection sampling: A technique where multiple outputs are generated and only those meeting a specific success or similarity criterion are kept for training

GAIA: General AI Assistants—a comprehensive benchmark designed to assess general intelligence and multi-step reasoning capabilities of AI agents

Pass@1: A metric measuring the percentage of times the agent's first attempt successfully solves the task

Pass@3: A metric measuring the percentage of times the agent successfully solves the task within three attempts

SFT: Supervised Fine-Tuning—training a model on labeled examples to follow specific instructions or generate specific formats

Playwright: An open-source automation library used to control web browsers, utilized here by the Web Agent

Accessibility tree: A text-based structural representation of a webpage's elements, used as input for the agent to understand the page without visual processing

Trajectory sampling: The process of recording the step-by-step actions and observations an agent takes to solve a task, used as training data

LangChain: A framework for developing applications powered by language models, used here for similarity-based matching during rejection sampling