When teams budget for AI agents, they think about LLM API costs. But the LLM call is often less than half the total cost of running an agent in production.
The Full Cost Stack
1. LLM API Calls (40-60% of Total)
The obvious one. But even here, teams underestimate. A single user query might trigger 5-15 LLM calls as the agent reasons, plans, executes tools, and synthesizes results. Each call has both input and output token costs, and input tokens are often 3-5x the output.
2. Embedding Costs (10-20%)
If your agents use RAG (retrieval-augmented generation), every document chunk needs an embedding. At scale, embedding costs add up. OpenAI's text-embedding-3-small is cheap per call but expensive at millions of documents.
3. Compute Sandboxes (10-25%)
Agents that write and execute code need sandboxed environments. E2B charges per second of compute time. A code generation agent that runs 30-second test suites on every iteration can cost more in compute than in LLM calls.
4. Tool Call APIs (5-15%)
Web search, browser automation, database queries, file operations. Each tool call has a cost. Browserbase charges for page loads. Search APIs charge per query. These add up fast when an agent makes 20+ tool calls per task.
5. Retries and Error Handling (5-15%)
Rate limits, timeouts, malformed responses, and hallucinated tool calls all trigger retries. A well-designed retry policy might add 10% to costs. A poor one can 10x them.
6. Infrastructure Overhead (5-10%)
Servers, databases, message queues, monitoring, logging. These are real costs that scale with agent volume.
Tracking the Full Picture
AgentBurn tracks all of these — not just LLM calls. Every cost event can specify any provider and operation type. Send your E2B compute costs, your Browserbase page loads, your vector DB queries. One dashboard for everything.
The first step to optimization is visibility. You can't reduce what you can't see.