Problem Description
I'm running an agent using OpenAI's GPT-4o model along with a Google Search tool. During execution, my agent may perform multiple searches, which results in multiple calls to the language model.
While debugging, I can see the input/output tokens for each step. However, at the end of the process, when I try to display the total input and output tokens used, it does not include all previous tokens that the agent has consumed across multiple steps.
Current Implementation
I'm retrieving metrics using:
response = self.agent.run(message=message, show_full_reasoning=self.debug)
print("METRICS: ", response.metrics)However, the response.metrics only reflects the most recent step and does not accumulate token usage over the entire process.
Proposed Solution
It would be helpful to have a session-wide history of token usage that keeps track of all intermediate calls made by the agent.
A few possible solutions:
- Implement a cumulative metrics tracker at the agent/session level that aggregates tokens across all calls.
- Extend
SessionMetrics to persist historical token data so that the total cost can be estimated accurately.
- Provide an API method like
agent.get_total_metrics() that returns the sum of all tokens consumed during an agent's lifecycle.
This would allow users to easily compute the total fees incurred when running multi-step agents.
Problem Description
I'm running an agent using OpenAI's GPT-4o model along with a Google Search tool. During execution, my agent may perform multiple searches, which results in multiple calls to the language model.
While debugging, I can see the input/output tokens for each step. However, at the end of the process, when I try to display the total input and output tokens used, it does not include all previous tokens that the agent has consumed across multiple steps.
Current Implementation
I'm retrieving metrics using:
However, the
response.metricsonly reflects the most recent step and does not accumulate token usage over the entire process.Proposed Solution
It would be helpful to have a session-wide history of token usage that keeps track of all intermediate calls made by the agent.
A few possible solutions:
SessionMetricsto persist historical token data so that the total cost can be estimated accurately.agent.get_total_metrics()that returns the sum of all tokens consumed during an agent's lifecycle.This would allow users to easily compute the total fees incurred when running multi-step agents.