JFrog measured 847 log lines to find 9, and that ratio is now a budget line

Yoav Landman, co-founder and chief technology officer of JFrog, told Lets Data Science that in one measured test a coding agent was handed an 847-line log when 9 lines were what it needed to diagnose the failure ^{[1] [2]}. That ratio, not the price of a token, is where his account puts the cost of agentic development, and it turns context hygiene into something a finance team can audit ^{[2] [13]}.

The backdrop is not contested. Per-token prices have fallen roughly 98 percent, with GPT-4-equivalent performance dropping from about $20 per million tokens in late 2022 to around $0.40, a decline of about 50 times ^{[3] [4]}. Over the same period enterprise AI bills rose an estimated 320 percent, with average annual budgets going from $1.2 million in 2024 to $7 million in 2026, roughly 5.8 times, and per-developer consumption up about 18.6 times in nine months, according to reporting by The Next Web citing Jellyfish ^{[5] [6]}. Uber exhausted its entire 2026 AI coding budget by April ^[7]. Microsoft revoked its developers' Claude Code licences six months after granting them, with individual engineers spending between $500 and $2,000 a month on tokens ^[8].

Landman's explanation of the mechanism is mechanical. "Because these agents operate in continuous loops, they retain prior conversation history to preserve their reasoning chain, meaning any terminal noise compounds with every single turn," he told Lets Data Science ^[9]. Run an npm install or a docker build inside a loop and the window fills with progress bars and status updates that you pay for again on every subsequent step ^[10]. He puts agentic coding at a thousand times the token consumption of standard code chat, a company figure the publication said it could not independently verify ^[11].

The product answer has two parts, by his description. JFrog Boost sits locally as a wrapper applying dynamic filters that strip progress bars, ANSI noise and redundant chat logs, with custom TOML filters available for teams whose internal tooling emits its own formats ^{[12] [14]}. For codebase search, a component called BoostGraph intercepts search results and prioritises the most relevant files and snippets rather than letting the agent pull in loosely related code ^[15].

The savings numbers are company-reported: 30 to 40 percent less token waste per CLI command in typical workflows, reaching 80 percent in cases like the 847-line log, with the answer stating plainly that 80 percent is the ceiling and not the average ^[13]. Applied to Microsoft's disclosed per-engineer range, a 30 to 40 percent reduction would be worth $150 to $800 a month per seat, if spend scales with tokens ^[16].

The detail that matters more than the percentage is the escape hatch. Landman says agents "can easily bypass filters to retrieve full raw outputs, or automatically disable filters that prove too aggressive" ^[17]. That is the difference between a slow correct answer and a fast wrong one, because a filter that removes something load-bearing produces confident failure rather than an error. He calls the emerging layer "Agentic Token Efficiency", a phrase from a vendor selling into it, and argues that "more tokens do not equal better output" ^{[18] [19]}.

Worth watching: whether the per-command reduction shows up on invoices rather than in benchmark decks, and whether bypass events are logged, since the rate at which agents reach past a filter is the only honest measure of how aggressive it was. The 847-to-9 count is also reproducible in-house. Any team can measure the signal share of its own build output before buying anything to fix it.