Build Proactive Personalized Assistant with AutoGen & Gemini 2.5 Pro
Inspired by Apple's shift towards a personalized, Gemini-powered Siri, this challenge tasks you with building a sophisticated multi-agent system using Microsoft's AutoGen framework. The goal is to create a proactive digital assistant that anticipates user needs, learns from interactions, and leverages dynamic tool use to provide personalized assistance in real-time. This system should be capable of understanding complex user contexts, synthesizing information from various sources, and initiating relevant actions without explicit prompting, mimicking a truly intelligent personal assistant.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a graph-based multi-agent system using AutoGen for a proactive, personalized digital assistant powered by Gemini 2.5 Pro with real-time voice and memory.
How work is evaluated
The evaluation will assess the system's ability to proactively assist users, correctly interpret complex requests, personalize responses based on historical data, and demonstrate robust multi-agent coordination.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Proactive Relevance
Suggestion is contextually relevant and not a generic chatbot response.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Tool Execution Success
All necessary tools are identified and simulated execution is successful.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Multi-Agent Coordination
Dialogue trace demonstrates clear, sequential, and logical agent interactions.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Personalization Score
How well the agent's response reflects learned user preferences and history. • target: 0.8 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Response Latency
Average time taken to generate a full response (simulated). • target: 3000 • range: 0-10000
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master AutoGen's conversational programming paradigm for defining complex multi-agent workflows and communication protocols.
Implement advanced prompt engineering techniques for Gemini 2.5 Pro to enable proactive, context-aware reasoning and response generation.
Integrate Deepgram's real-time Speech-to-Text and Text-to-Speech APIs for seamless voice interaction within the agent system.
Design and manage long-term personalized memory using Qdrant vector database for storing and retrieving user preferences, history, and context.
Build a dynamic tool invocation mechanism within AutoGen agents, allowing them to autonomously select and execute relevant actions or retrieve information via external APIs served by AI21 Studio's inference endpoints.
Orchestrate agent roles, such as a 'Context Analyst,' 'Action Planner,' and 'Information Retriever,' to collaborate effectively on complex user requests and proactive suggestions.
Deploy and manage multiple AI models using AI21 Studio's platform for efficient serving and routing of specialized tasks (e.g., summarization, entity extraction).
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
The evaluation will assess the system's ability to proactively assist users, correctly interpret complex requests, personalize responses based on historical data, and demonstrate robust multi-agent coordination.
Challenge input
{'user_history': [{'text': '...'}, {'text': '...'}], 'current_context': {'time': '...', 'location': '...'}}
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{'proactive_suggestion': '...', 'reasoning': '...'}
- Suggestion is contextually relevant and not a generic chatbot response.
- All necessary tools are identified and simulated execution is successful.
- Dialogue trace demonstrates clear, sequential, and logical agent interactions.
- Personalization Score target: 0.8
- 2 public reference cases
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Proactive Relevance · Weight 1
- Tool Execution Success · Weight 1
- Multi-Agent Coordination · Weight 1
- Personalization Score · Weight 1
- Response Latency · Weight 1
Metrics
- Personalization Score · Target 0.8
- Response Latency · Target 3000
Reference cases
- 2 public reference cases
Workflow tasks
- Task 1
Input: {'user_history': [{'text': '...'}, {'text': '...'}], 'current_context': {'time': '...', 'location': '...'}}
Output: {'proactive_suggestion': '...', 'reasoning': '...'}
Execution harness: Python
- Task 2
Input: {'user_query': '...', 'tools_available': ['tool_name', ...]}
Output: {'final_response': '...', 'tools_used': ['tool_name', ...], 'agent_dialogue_trace': [...]}
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.