I work on Claude Code. You've got the basic mechanism right, but to be super clear, this only applies to background agents that are actively working. They get a summary call about every 30 seconds, and idle background agents don't. These summaries are used for both you and Claude Code to track what each background agent is working on, without reading transcripts.
Your 16.4% sounds like it was for a team running mostly background agents, and doesn't seem like a typical session. (Feel free to send me the session ID at edwinarbus@anthropic.com and we can dig further.)
Last week, across Claude Code users who send us telemetry, agent_summary was about 2% of cost at API list prices and just under 4% of requests. In the terminal, where nearly all of these calls happen, it was about 6% of cost.
That said, we agree that's a lot for a status line. We just rolled out a fix for the most expensive case, summaries that miss the prompt cache (2.1.311+), and right now we're working on a much cheaper mode that sends short notes to a smaller model instead of the full transcript.
That helps put the 16 percent in context. I think it would be better to add a real setting to turn agent summaries off... in the current binary, both the sdk option and the env var can only switch them on
So are we just holding it wrong, or is the distinct absence of any mechanism to tune/turn it off for our own good?
Also why you're so dismissive with the 2/4% figure while completely ignoring the TFA analysis and the other examples in their linked github issue?
Seriously, you're adding a feature that silently burns 15-25% tokens (that is burning money for API clients while hiding that fact), take away everything that could be used to mitigate it, and your response is basically, "eh, your numbers are unimportant and ours say it's negligible but since you complain we'll make it a bit cheaper".
Seriously, that looks so intentionally customer-hostile and revenue-friendly, but reading the room no one is surprised.
You're telling us there are frequent scheduled summary calls to agents that burns tokens like this is common knowledge. Is it? And is this documented somewhere?
Perhaps this is terminology confusion, but the official docs explicitly state that status lines "[Run] locally and [do] not consume API tokens." [0] _If_ that documentation is talking about the same thing that TFA is talking about, then I'd say that this is very much not common knowledge.
I also wonder if the fix that they "just rolled out" is actually activating the 'short' mode that TFA mentions...
Your 16.4% sounds like it was for a team running mostly background agents, and doesn't seem like a typical session. (Feel free to send me the session ID at edwinarbus@anthropic.com and we can dig further.)
Last week, across Claude Code users who send us telemetry, agent_summary was about 2% of cost at API list prices and just under 4% of requests. In the terminal, where nearly all of these calls happen, it was about 6% of cost.
That said, we agree that's a lot for a status line. We just rolled out a fix for the most expensive case, summaries that miss the prompt cache (2.1.311+), and right now we're working on a much cheaper mode that sends short notes to a smaller model instead of the full transcript.