Profiling the agent loop: Jev handles about 71% of decision steps
Profiling an agent loop as slow code shows most steps are small forks like ignore, route, score or is it safe to run, with Jev taking about 71% of steps while the LLM writes in about 14% yet consumes most of the tokens.
THE MOST COMMON THING YOUR AGENT DECIDES IS TO DO NOTHING profile an agent loop the way you'd profile slow code and the picture flips. the tall bars aren't writing they're tiny forks: ignore this, route that, score this, is it safe to run in this schematic jev takes ~71% of the steps. the LLM writes in ~14%. code acts in the rest. the LLM still eats most of the tokens. it just shows up way less often than you'd think now look at the pareto in the bottom right. "ignore" is the biggest bar. then route, score, reply. "human" is small on purpose that's the job people pay frontier prices for. a big model reads an email, writes a paragraph about it, and lands on "skip". jev doesn't write the paragraph. state and a typed question go in, a choice, score or yes/no with a probability comes out. "phishing?" -> YES in 184ms. "next worker?" -> write at 0.91. and when it's unsure it says so. safe_to_run 0.62 is under the 0.85 gate, so the job goes to a human instead of shipping auto-post to X in this setup: never. drafts wait for me to hit post before you shop for a cheaper model, count your forks first. a lot of that bill is your agent deciding to ignore things
This post quotes another post: x.com/asteri_eth/status/2104901781725598103