UCSD Agent testing

Notes

Langchain's Deep Agents vs our own Implementation

Deep Agent provides:

• perfect Skills integration (just like Claude)
• planning/task-decomposition
• long-term memory
• context management
• sub-agent spawning
• customizable backend (state/file-based)
• human-in-the-loop support
• self-improving/modifying instructions/prompting

Deep Agent Architecure: Todolist, Fileystem, and subagent are provided customizable middleware.docs

They inject their default messages onto your system prompt, and you cannot change them (which is a giant pain)

Tools vs Skills vs MCP server:

Tools: Built in and supported by Langchain. The middleware for this is perfect and completely handled.

Skills: For langchain/graph not Deep Agent: We'd have to build our own middleware. It'd be confusing for us to differentiate what we turn into a tool vs skill. A decent amount of intellectual overhead and would probably lead to bloat. I made an early attempt

MCP Server: Can have multiple MCP servers. Use langchain mcp adapter and it allows langchain to just see/use/call them as regular langchain tools.

It seems like an agent can't add/initialize new MCP servers during runtime, as the MCP servers are loaded and transformed into tools and then passed to the agent during the agent's intitialization.

Relevant Langchain Functionality

• Can force structured output
• Prompt chaining/states for predetermined workflows. Can also add parallel nodes/states
• Routing for smart routing of queries/text via structured output
• Orchestrator+worker can force orchestrator to follow a strict planning schema and then can pass subsections onto specific workers
• Evaluator+optimizer by using structured output + nodes

Useful docs on their multi-agent techniques

Current Architecture

Now it's just the default Deep Agent architecture + new tools + skills.

Added tools:

• curl: As named

Manually added skills:

• get-weather: a claude skill which I copied/downloaded
• skill-creator: as named, taken from link (made by langchain)

Setup

Conda environment for clean local dev environments

conda create -n "agents_ucsd" python==3.11
conda activate agents_ucsd
pip install -r requirements.txt
# ask the agent to construct a skill from prompts/download_video_input.txt 
python main.py