High input token usage

[edvanr]

I've been playing around with Strands Agents and loving it. My use case is around a meeting assistant, and it is helping me with my meetings and calendar data.

I've been testing it out for a few days and just realised my AWS bill for the model has jumped up over $50.

I used some of the metrics to find out token usage and it looks like the majority of it is coming from input tokens.

For a query like "What meeting do I have today?" I'm seeing about 45k tokens being used, which costs around 14c. Does that sound right?

Are there ways to optimise the amount of input that Strands is generating?

[awsarron]

Input tokens are made up of:

1. System prompt
2. Tool specifications (name, description, input schema)
3. Agent state - user messages, assistant messages, tool calls, tool results

When running agents, tokens for each cycle can be seen in the AgentResult object (https://strandsagents.com/latest/user-guide/observability-evaluation/metrics/). For example:

from strands import Agent
from strands_tools import current_time

agent = Agent(tools=[current_time])
agent_result = agent("What time is it?")

print("\n\n")
print(agent_result.metrics.get_summary())

Additionally, Strands integrates with OpenTelemetry traces: https://strandsagents.com/latest/user-guide/observability-evaluation/traces/.

Can you share a full code sample and your input messages to the agent that reproduces this issue?