Jev floods production, Theo says the plugin is broken
Vercel says Jev reached about 13% of its AI Gateway teams on day one, the fastest adoption it has recorded. Then Theo took apart the compaction plugin everyone was installing, and a developer who benchmarked it agreed with him.
The big story
Greg Isenberg says Jev sorted 1,700 emails for 18 cents, Tony Dinh runs it at $0.005 a video
Jev is showing up in shipped products and people are posting the bills. Jon Yongfook put Jev-powered field mapping live in Bannerbear, matching template fields like `name` to source fields like `full_name` in one click. Tony Dinh built an open-source Chrome extension that detects and skips YouTube sponsor segments in real time for about $0.005 per video. Greg Isenberg says it sorted 1,700 emails for 18 cents, and Yohei says he still has not spent $3 on it. Guillermo Rauch says adoption is "shocking" but posted no numbers, so treat that as a vibe, not data.
1,700 emails for 18 cents, instantly.
Jev is fast and cheap but is not very smart. No reasoning, no thinking.
how have i still not even spent $3 on this thing
Real numbers
Lemkin says agents closed $2.7M in leads reps never touched
Jason Lemkin published the breakdown behind his claim that SaaStr doubled revenue with agents: three humans, 20+ agents in daily production, revenue moving from -19% year over year to +47%. He says a sales agent worked about 1,000 warm leads sitting untouched in Salesforce, got a 72% open rate and closed $2.7M, and that two other agents, an AI VP of Marketing and an AI VP of Customer Success, cost $257 a month combined. He also says one of them found two customers still being billed $300 a month for a product killed six years ago. His full write-up includes the unglamorous part: daily QA.
The takeaway: do not put AI on your A leads. Put it on the B leads your humans will never follow up with. That is where the real revenue is.
And the part nobody puts in the tweet thread: vibe-coded apps need daily maintenance. Agents drift, data changes, models update, integrations flake. We QA ours every day.
Arguments and hot takes
Theo calls the Jev compaction plugin "a terrible compaction strategy"
Theo's argument: compaction (shrinking an agent's conversation history so it keeps working) is not a filter, it should run sparingly rather than constantly, and Jev decides per tool call without knowing what the tool call returned. He also says frontier models from OpenAI, Anthropic, XAI and Google ship reasoning as encrypted payloads that Jev cannot see, and that editing early history forces cache rewrites for everything after it. On cost, he says cache writes often run over 60% of his total LLM spend in his own use of Claude Code and Codex. Replying to @tamarajtran's point that tool calls are matched by id, Theo said that is no longer true for Anthropic models. When @moinerus benchmarked the plugin, Theo thanked them and said the results went his way.
This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.
Cache writes are the biggest cost by far for agents. I often see cache write costs go over 60% of my total LLM spend in my personal use of Claude Code and Codex.
Thank you for benching it! Would have been awesome if I was wrong here. Sad to see I wasn't.
Quick hits
Marc Lou says a sponsor spent $68,000 on his body and $30,000 in credits
Marc Lou says Higgsfield spent $68,000 on his tattoos and gave him $30,000 in API credits, which he turned into a puzzle game giving away 30 prizes of $1,000 in credits. He later reported 1,502 people found the first clue and 912 finished, and separately said he has never taken money to say good things about a product.
Yongfook loses a multi-year customer to a vibe-coded replacement
Jon Yongfook says a long-term customer, itself a SaaS company, cancelled, and he assumes they built their own version, with the lesson don't target developers. levelsio replied that his own custom stuff is more custom and breaks less than paid SaaS, which is the churn reason in a sentence.
Tibo says Squad crossed $10k MRR after doubling in September
Tibo reports Squad reached $10k MRR and that revenue doubled in September alone, with the pitch being that you plug your own Anthropic and OpenAI subscriptions in and stack them. He credits built-in scrapers, email and integrations so agents can do outreach out of the box.
Claude Code now reads AGENTS.md when there is no CLAUDE.md
Thariq announced that from version 2.1.277 Claude Code will fall back to AGENTS.md when a folder has no CLAUDE.md, toggleable in config. Simon Willison said it means he can stop creating CLAUDE.md files that just point at AGENTS.md.