agent-sofia

| Creating agents for better health care understanding.

By:

• Susan Kraemer, PhD
• Cameron Guthrie
• Patrick Cavins, PhD

Background

An agentic application can have multiple users and each user may have multiple sessions with the application. To manage these sessions and events, ADK offers a Session Manager and Runner. Created as part of the 5-Day AI Agents Intensive Course with Google

SessionService: The storage layer

Manages creation, storage, and retrieval of session data Different implementations for different needs (memory, database, cloud)

Runner: The orchestration layer

Manages the flow of information between user and agent Automatically maintains conversation history Handles the Context Engineering behind the scenes

Project structure

1. capstone-project-with-mcp-server.ipynb

• This is an end-to-end jupyter notebook that uses an MCP service with a single agent to answer medical queries
• Example log, demo_logs_from_query-script.txt
• Kaggle notebook, https://www.kaggle.com/code/camreon/team-seattle-capstone-project

2. multi-agent-and-custom-tool-script.py

• This is python script which creates multiple agents. One agents uses the Pubmed MCP, and the other agent uses a custom tool called get_dict_from_query which return a structure python dictionary using the format {'PMID': list[int]}. Both agents are sub-agents inside of root_agent which coordinates their invocation.
• Example log, `demo_logs_with_mcp_server.txt

3. README.md

• The explainer, hopefully.

4. b4-soup-edit

• workin progress python file for parsing a full text html publication

References

PUBMED MCP

• GITHUB
• REDDIT

PUMBED CLIENT

• GITHUB

PUBMED API CLIENT IN PYTHON

• University of Alabama, Library

Searching Based on Publication Types

We can sort by publication type by adding AND into the search:

term=<searchQuery>+AND+filter[filterType]

[pt] specifies that the filter type is publication type. More filters can be found at PubMed Help.