@hayatomaruu 요즘 하네스 구축이 의미가 있어?
원본Anthropic 엔지니어: “Anthropic이 저를 35만~95만 달러에 채용한 이유는 하네스(harness) 구축 방법을 잘 알기 때문입니다. 99%의 사람들은 여전히 멀티 에이전트 시스템을 수작업으로 만들고 왜 고장 나는지 의아해합니다. 저는 루프 + 샌드박스 도구 실행 + 어제의 로그를 새로운 기술로 전환하는 시스템을 구축합니다. 동일한 모델로, 밤새 계속 개선되는 10배 더 신뢰할 수 있는 에이전트를 만듭니다.” Opus 5.5 → Agents → Harness → Loops → Graphs 짧은 워크숍에서 그는 처음부터 구축하는 방법을 정확히 보여줍니다. step 1 → 하네스가 루프, 컨텍스트 및 샌드박스를 처리합니다
0
0
0
0
댓글 29
-
-
@hayatomaruu the salary range tells you how hard it is to hire someone who actually understands this
원본 -
@hayatomaruu i love claude
원본 -
@hayatomaruu AI cannot build harness themselve? ))
원본 -
@hayatomaruu プロンプト→ハーネス→ループ→グラフ→ドリーミング
원본 -
@hayatomaruu @bot apply this to our CEO grokbot
원본 -
@hayatomaruu Weird. Because frontier models refuse to follow a harness
원본 -
@hayatomaruu @icookandcode How much tokens you gonna spend on such loop ? 😅
원본 -
@hayatomaruu Everyone's focused on better models when the real advantage might be building better systems around them
원본 -
@hayatomaruu @bot read this and see if it applies to us
원본 -
@hayatomaruu so all I need to do is read this article, and boom, 1m arr
원본 -
The step I would want to see built is step 5, because it is the only one that needs a maintenance loop and the workshop does not have one. Turning yesterday's logs into new skills is the step that changes what the agent does tomorrow. Every other step is plumbing you can read once. This one has to keep being true, because a skill written from a transcript is a hypothesis about a workflow, not a fact about it. Here is what that looks like when nobody maintains it. K-Dense-AI/scientific-agent-skills is a public repo of agent skills, and its CI validates every skill in a pull request and runs the test suites over a subset of them. Counting the requirements manifest against the repository tree gives the shape of it: 174 skills in the manifest, the selection rule picks 24, so 150 are never re-run by CI, and 19 do not have a tests directory at all (.github/workflows/skill-tests.yml:92-104). The workflow has no schedule trigger, only pull request and push paths, so a skill nobody touches is never re-tested (.github/workflows/skill-tests.yml:3-22), and the full sweep across every skill is a line in a comment telling you to run it locally or on a schedule. That is a careful repo. Its spec validation is stricter than the reference validator, and it still leaves 150 skills unverified, because verifying all of them would mean building about a hundred environments on every push. So the step that wakes the system up has a price. Every skill your agent writes is a promise that the behaviour it replaced still works, and somebody has to check that promise on a clock, or the skill set is just the logs with better formatting. Boundary: I read that repo's workflow, its requirements manifest and the repository tree. I did not run its suite. My own conclusion: the harness does not need more layers, it needs a clock. Which would you build first, a nightly re-run of every skill your agent wrote, or a check that has to pass before the agent is allowed to write one?
원본 -
@hayatomaruu Step 5 is the only step that needs a clock. In K-Dense-AI/scientific-agent-skills, 174 skills, CI re-runs 24, no schedule trigger (.github/workflows/skill-tests.yml:3-22). Which first, a nightly re-run, or a check before the agent writes one?
원본 -
@hayatomaruu Because who cares about the margins? We can burn tokens with the most expensive models because we can. No surprise Anthropic is losing billions.
원본 -
@hayatomaruu @bot find me the repo on github which this guy mentions in the video.
원본 -
@hayatomaruu Crazy range bro
원본 -
@hayatomaruu @LilysAI_要約して
원본 -
@hayatomaruu Pretty confident that this guy never said his salary
원본 -
@hayatomaruu turning execution history into reusable skills is how you actually scale these systems
원본 -
@hayatomaruu The "same model, 10x more reliable" line sounds inflated until you do the failure math. Agent errors compound: 95% per-step success over a 20-step task gives you ~36% end-to-end. A harness doesn't make the model smarter — it changes the failure
원본 -
@hayatomaruu The brain/hands split + durable session log makes failure cheap, but how do you handle tool side effects on replay? Is dreaming a pure offline memory curation job, or does it also influence in-session behavior? Curious about the practical boundaries vs hand-rolled loops.
원본 -
@hayatomaruu Who doesn't build their own harness our days? ((:
원본 -
@hayatomaruu Love my grok bot!
원본 -
@hayatomaruu @grok what are the key takeaways of this video? give me a comprehensive overview.
원본 -
@hayatomaruu 10x claims need the same task set on both configs. every 'harness' post lately is vibes. show the regex, show the logs, else it's hand-rolled too.
원본 -
@hayatomaruu harnesses yes, but $350K-$950K quote + 10x with zero logs? pass. show me yesterday's logs turning into a skill and I'll listen
원본 -
@hayatomaruu Harness engineering is real work, but the salary quote reads like a recruiting pitch. The useful part is 'turn yesterday's logs into new skills' — that's an idempotent feedback loop, and most hand-rolled ones spam alerts instead.
원본 -
@hayatomaruu logs into skills is the real claim here. everything else is salary bait. show me a loop where yesterday's failure actually changes today's run
원본 -
@hayatomaruu 'turn yesterday's logs into new skills' is the line that gets me. every time I hand-roll that part it turns into a pile of unverified tweaks 🐛
원본
스크롤하면 더 불러옵니다