Skip to content
AI Training

Inspecting what the agent sends

Customizing agents · topic customizing-agents/inspecting-the-agent

What an agent sends to the model is hidden by default. This topic covers watching how full the context is while working, reading the record a harness keeps of each session, and logging the exact requests the agent sends to the model provider, along with why each of these records holds sensitive data.

Concepts

Context meter
A live view of how much of the context window a session uses. A status line can show the share in use and the tokens so far as the session runs, and a context command breaks the use down by category, such as instructions, tools, memory and messages. Watching it shows when a session is getting full and which tools or files are expensive, so the learner can compact or start fresh before the agent starts dropping instructions. glossary
Session record
The file in which a harness saves a session, with every message, tool call and tool result, so the session can be resumed later. It is the most complete account of what the agent saw and did, and the place to find where a run went wrong. It is also plain text on disk, kept for a set number of days, and it holds whatever passed through a tool, including file contents and any secret the agent printed. glossary
Request logging
Recording the exact requests an agent sends to the model provider and the responses it gets back. The agent can be pointed at a gateway between it and the provider that logs every request, or a monitoring export can be set to include full request bodies, which it leaves out by default. The log shows what each tool description, file and instruction costs. It holds everything the agent read, so it stays on systems the team controls. glossary

Lessons

Your reference

Each lesson above adds its takeaways and its example here once you finish it. Your reference lists every lesson you have finished.

Seeing what the agent sends

Unlocks when you finish Seeing what the agent sends.

Sources