正在加载视频...
视频加载失败
this is pure f*cking gold for anyone running coding agents Jev founder Diogo Amogo wrote a PDF on building a Jev harness what their tests show: > up to 200x faster > up to 400x cheaper the model hasn't been the bottleneck for a while speed and cost both... show more
33 条评论
400x cheaper numbers always make me squint. cheaper vs what baseline tho, someone running opus on max effort for a typo fix?
The biggest harness win I've seen running client agents isn't caching, it's trimming tool output. A test run that dumps full logs into context can cost more than the fix itself. Tail or summarize tool results before the model sees them and the bill drops without touching the model.
the 200x/400x numbers only hold inside their harness setup, but the framing is right. model stopped being the bottleneck a while back — watcher loop + durable state is where the wins live.
@grok is it true?
'200x faster 400x cheaper' is a benchmark claim, not a receipt. which workload, baseline harness, same task set? would read it if the numbers ship with the test
200x faster and 400x cheaper from the harness is the receipt that stuck. i still burn quota on a bigger model before i rewrite the wrapper
cost and speed both coming from the harness means everyone’s tuning the wrong thing
gree the harness is where the gains are now. Same model, different tool loop, context strategy and verification step, and you get a very different agent. The wrapper shapes behavior more than benchmarks suggest.
jev
The speed and cost advantages for agentic workflows are massive. We actually went deeper on this here:
200x is always the harness. i've rebuilt the same loop five times this year.
Jev 这个词这几天刷到好几次了…200x 400x 这种数字先信一半，harness 比模型重要这句我倒是认，瓶颈早就在工具执行和状态管理那了
harness tuning cut my arm sim render time from 40ms to 2ms
Harness matters, but the model gap showed up for us anyway. Each in its own harness, same 6 repos: one fixed 24 of 24, the other 20, and every miss was one repo. What do you measure in your harness, cost per fix or pass rate?
I don’t understand the hype behind Jev its just a bert but less accurate
"the model hasn't been the bottleneck" is the whole thing. 200x faster / 400x cheaper is a harness number, and I'd rather see the traces than the multiplier.
harness才是真差距，这点早该有人喊了。但200x/400x这种数字，基线是啥得写清楚，不然又是营销话术。模型从来不是瓶颈，工程才是。
cutting unnecessary model calls is a compelling reason to work on the harness. swapping models wouldn't remove that work.
200x faster / 400x cheaper is a big claim without the test setup. did they hold the model fixed and only swap the harness? that's the version i'd want to reproduce.
model早不是瓶颈了，harness才是。200x这个数得看baseline怎么设的，但方向没毛病，做agent的都在踩这个坑
200x速い400x安い、の数字は条件が書かれてないと意味が薄い。watcherもいないcron運用なら、速度より先に失敗の検知とロールバックを確認したい
别吹牛逼了
That speed boost is wild. If Jev's harness really cuts costs that much, it could change how teams approach building and scaling coding agents. Everyone's gonna want to optimize after seeing these results.
200x/400x这数字先信一半，但'瓶颈早就不在模型了'这句讲真扎心。harness才是现在真正拉开差距的地方。
200x faster / 400x cheaper is great but always net of fees the harness buys you. noticed most of my edge came from killing wasted calls, not the model swap. fwiw
200x faster and 400x cheaper in one pdf is a big claim, i'd want to see the test setup first
The promotion angle is smart because it connects engineering improvements directly to business outcomes.
Why would any care about Jev @grok
wow, the pdf from Jev’s founder is actually gold
A harness is the workshop around the engine: caching, retries, and focused tests can cut wasted runs even when the model stays the same.
200x / 400x numbers are doing a lot of work in that screenshot. has anyone actually reproduced the harness benchmark, or is it one cherry-picked task? genuinely asking, not dunking.
200x / 400x against what baseline tho. default harness, or their own previous config? numbers like that usually hide the setup. got a link to the actual tests?
Swapping models feels like progress, granted; but the harness is where all the actual leverage lives. When an agent spins out, people blame the weights; usually it is prompt drift, bad compaction boundaries, or broken tool loops.
![this harness stack is f*cking gold TypeSafe AI + LangChain finally dropped the layer that makes agents affordable at company scale: jev running inside the langchain agent loop as middleware the problem every big team hits: each fork in the loop is another full LLM call. at 10 agents that's a bill. at 10,000 it's the budget > routing before: every task goes to the frontier model now: jev picks the least costly model that can finish it. lookups go fast, architecture goes powerful > guardrails before: the "is this command dangerous?" check sat in the closed parts of claude code, codex, cursor now: AutoModeMiddleware(tools=["bash"]). one line, any agent, blocked before the tool runs > decisions before: one LLM call per question now: every question about the same state in one request, answered in parallel > audit before: "the model decided" now: the answer, the probability and the confidence stay in agent state > adoption no rewrite. pip install langchain-typesafe and drop it into the agent you already run TypeSafe reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification. the LLM still reasons and writes. jev takes the thousands of small calls in between already live: browserbase agents for fractions of a cent, a live trading agent, email triage at scale full loop mapped below: 7 agents, 4 jev gates per step count the forks in your agent loop tonigh](https://image.24vids.com/tw-2104876056893104461/amplify_video_thumb/2104866254146203648/img/J8LnvuFvwbGYaJ7N.jpg)
