How can I lower the costs of running my AI agent?

There are several levers that affect how much it costs to run your AI agent. Here's what you can control:

1. Frequency of Use

How often your agent runs is a major factor. Triggering it every few minutes versus once an hour can lead to a substantial difference in monthly cost.

2. Size of the Context

Feeding the agent too much data — especially unnecessary context — quickly adds up. Trimming the context to only include what's essential is one of the most effective ways to cut costs without sacrificing quality.

3. Where You Store Data

This is an often-overlooked but powerful design choice. If you embed data directly into the agent's instructions, that data is included in every single run — even when it's not needed. By contrast, if you store supporting information in documents or tools that the agent can selectively retrieve, that data is only included when relevant. This can dramatically reduce the tokens used and save credit over time.

4. Complexity of Instructions

Recursive or multi-step instructions — where the agent performs several tasks or calls itself — can drive up usage significantly. Simplifying the agent's logic and reducing unnecessary loops can help manage costs effectively.

5. Language Model Selection

Currently, we use Claude 4 as our default model because we've found it provides the best balance of performance versus cost and is the easiest to get agents working reliably with. However, model choice can significantly impact costs.

Current Approach

We prioritize getting agents to work effectively first, then optimize costs later.

Future Options

We're exploring the possibility for users to choose different models based on their needs:

• Cheaper models for simpler tasks
• More expensive models for complex requirements
• Alternative models based on specific use cases

Cost Optimization Support

We're happy to help with cost optimization on a case-by-case basis, but haven't yet enabled universal model selection for all users.

Your Feedback is Welcome

This is an evolving area - let us know your thoughts and requirements around model selection and cost optimization.