|
|
|
|
|
|
by eugeneonai
13 days ago
|
|
|
Looks nice. The README pitches model-per-task picking but doesn't say
much about context management. In coding-agent loops the full system
prompt + tool specs re-send on every step — a 30-step task pays the
input cost 30x. Prompt-cache headers catch the static prefix, but the
per-step diff (file diffs, observation tokens) isn't cached, and that's
often most of the input. Auto-summarizing older trajectory into a state
vector saved 40-60% input tokens in workloads I've looked at — could be
a useful daemon-side concern since users won't reach into each agent's
internals. |
|
|