容易造成超出模型token的问题，能不能提供改进adk上下文相关的问题

[tangguibiao]

I'm developing a multi-agent system similar to Meta's "Manus," and encountering issues with the model's distorted tool-calling capabilities, primarily related to context. We eagerly hope the official team can provide improvements to the ADK's context-related methods. Thank you

[wuliang229]

Can you be more specific about your question? For example, what is the setup and what models/tools are you using?

[hangfei]

Move to discussion/question as it's a question.

[hangfei]

@adk-bot to help answer.

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: You can manage token usage by controlling agent history, using session state for inter-agent communication, creating specialized stateless agents, and leveraging context caching. ADK also provides tools to track token usage for precise optimization.

---

Hi @tangguibiao,

Thank you for your feedback. Managing context length and token consumption is indeed a critical aspect of building complex, multi-agent systems. The ADK provides several mechanisms to help you control and optimize this. Here are some strategies that might help address the issues you're encountering.

1. Control Agent History

By default, LlmAgent includes the entire conversation history in the prompt sent to the model. You can control this behavior using the include_contents parameter in the agent's constructor. For agents that perform specific tasks and do not need full historical context, you can set this to 'none'. This will create a "stateless" agent that only receives the current turn's content, significantly reducing the number of tokens sent to the model. [1]

# This agent will not receive the full conversation history
stateless_agent = LlmAgent(
    model="gemini-1.5-flash-001",
    include_contents="none",
    # ... other parameters
)

2. Use Session State for Inter-Agent Communication

In a multi-agent workflow, instead of passing large amounts of text between agents (which adds to the context window), you can use the shared session.state. One agent can perform a task and write its result to session.state, and a subsequent agent can read from it. This is a more token-efficient way for agents to communicate and share data within a single session. [2]

For example, you can use the output_key property on an agent to automatically save its output to a specific key in session.state. [2]

3. Leverage Context Caching

ADK includes a context caching feature that can help reduce token usage on repeated requests. By enabling ContextCacheConfig, you can cache parts of the context that are frequently reused. This is particularly useful in scenarios where you might be querying a large document repeatedly. There is a sample that demonstrates how to use this feature effectively. [3]

4. Track Token Usage

To better understand and debug where your tokens are being consumed, you can track token usage for each turn. The ADK provides usage_metadata in the event stream, which includes prompt_token_count, candidates_token_count, and total_token_count. By monitoring these values, you can pinpoint which agents or turns are contributing the most to the token count and optimize them. [4]

You can implement a callback to access this data:

async def handle_events(events):
    async for event in events:
        if event.type == "usage_metadata":
            print(f"Total tokens used: {event.usage_metadata.total_token_count}")

We hope these suggestions help you better manage context and token usage in your multi-agent system. We are continuously working on improving context management features and appreciate your valuable feedback.

[1] publicly accessible HTTPS URL of the document: https://google.github.io/adk-docs/agents/llm-agents/
[2] publicly accessible HTTPS URL of the document: https://google.github.io/adk-docs/agents/multi-agents/
[3] publicly accessible HTTPS URL of the document: https://github.com/google/adk-python/blob/main/contributing/samples/cache_analysis/README.md
[4] publicly accessible HTTPS URL of the document: https://github.com/google/adk-python/blob/main/contributing/samples/token_usage/main.py