Agent efficiency
Context carry
An agent step doesn't send a question. It sends the whole session so far: the system prompt, every tool schema, and every tool result since the start. A file read at step 3 is read again at every later step. Move the sliders and watch the pile grow, then turn on clearing old tool output.
My main profile, Oct 20–28: about 49 model calls per session.
A command's output or a file read. Sizes vary step to step around this average.
System prompt + tool schemas, sent every call. Mine: 12.4K + 8.6K.
Hermes' compression.proactive_prune_tokens: past the trigger, results over 8,000
characters outside the last 6 messages become one-line summaries. Off by default.
- Read this session
- –
- Read per token written
- –
- Old tool output share
- –
- Clearing saves
- –
The agent writes ~150 tokens per step: a tool call or an answer.
Turn on clearing to compare.
Measured on my own agent, Oct 20–28: 131,100 tokens read per call, 50.8% of it old tool output, each tool-output token re-read 54×. The full audit →