The Problem
Teams building AI agents that browse the web face a recurring integration burden: wiring together search APIs, browser automation, model providers, and tool execution. Each integration has its own SDK, auth model, and failure modes. agentic-internet attempts to consolidate these into a single agent framework with a unified interface.
What This Does
agentic-internet is a Python framework built on HuggingFace's smolagents that provides an InternetAgent class for autonomous web research and interaction. The package is organized into agents/ (agent implementations), tools/ (browser automation, web search, code execution, MCP integration), and config/ (settings). A CLI in agenticinternet/cli.py exposes commands, and main.py serves as a secondary entry point.
The framework supports multi-model orchestration via searchorchestrator.py and multimodelserpapi.py, routing tasks to different models (OpenAI, Anthropic, OpenRouter) based on task type. It includes SerpAPI integration for multi-engine search and a Browser Use Cloud SDK wrapper for form filling and structured data extraction.
How To Use It
Setup: Install with uv sync (recommended) or pip install -e . per the README. A uv.lock file exists, so dependency resolution is reproducible under uv.
Configuration: Create a .env file in the project root. Required keys are documented in .env.example and include OPENAIAPIKEY, ANTHROPICAPIKEY, SERPAPIAPIKEY, BROWSERUSEAPIKEY, and OPENROUTERAPIKEY. All are optional; the agent degrades to available providers.
Running it:
CLI
uv run python -m agenticinternet.cli --help
Programmatic
from agenticinternet import InternetAgent agent = InternetAgent() result = agent.run("Search for the latest AI news and summarize the top 3 stories")
Real-World Use
A research team automating competitive intelligence gathering could use this to replace manual browser workflows. The orchestrated search example (agenticinternet/examples/orchestratedsearchexample.py) shows a pattern: run a query across multiple search engines via SerpAPI, validate results cross-engine, then have a specialized agent (news analyst, data analyzer) process the output. The MCP integration (agenticinternet/tools/mcpintegration.py) allows extending the agent with external MCP servers, which is useful for connecting to internal tools.
Code Health & Issues
Medium - No CI/CD pipeline: No .github/ or CI config detected. Nothing gates merges on test or lint passing. The 14 test files exist but run only if someone remembers to execute them. Low - No verified test execution: Tests are present (tests/testmcpintegration.py, testagent.py, testsimple_agent.py) but there's no evidence they pass or are maintained. The root-level test files suggest ad-hoc testing rather than a structured suite. Low - Dependency lockfile mismatch: pyproject.toml declares dependencies and uv.lock exists, but the analysis flags no lockfile. Verify uv.lock is current before relying on reproducible builds. Low - API key sprawl: Five optional API keys means the agent's behavior varies significantly based on what's configured. Error handling for missing keys is not visible from the structure.
The Bottom Line
agentic-internet is a functional prototype that demonstrates a reasonable architecture for multi-provider web agents. The multi-model orchestration and MCP integration are the most distinctive features. It's not production-hardened — no CI, unclear test status, and heavy reliance on third-party paid APIs. Use it for prototyping agent workflows or as a reference for building your own orchestration layer, not as a drop-in production dependency.