
Product Engineer @passionfrootme likes: product, search, recommendation systems, genAI, games prev.: drop out @ycombinator PM @jinaai_ Founder @ https://teapawt.org/t.co/NfaaWkcYNa
Germany
Joined June 2019
- Tweets61
- Following93
- Followers81
- Likes67
I spent some time to look into leviathan by @joshuagunnn, a BM25 index over your data for coding agents. It claims 245x fewer tokens than grep on database access. Unfortunately (Spoiler) it did not hold up in my tests :(
github.com/elstongun/leviath…
The grep baseline has no limit, and the harness counts tokens on its full output. The 96% grep accuracy in the README only looks at the first 30k chars. Uncapped, grep finds a relevant record in 200/200 questions vs leviathan's 198
1
17
Claude CoWork now has a built-in skill to create marketing campaigns.
But context matters! General AI like ChatGPT and Claude struggle to go beyond the obvious.
Our Zest Agent is built to put the right information in the right place, at the right time.
v1 ✅ what's next?
Creator-led GTM works. Scaling it hasn’t.
Until now.
🍋 Introducing Zest, the first AI agent for creator-led GTM.
From one prompt, Zest builds strategy, finds creators, runs outreach, manages campaigns, and turns results into clear next steps.
Modern GTM is creator-led. With Zest, it’s finally scalable.
agents -> MCP -> CLI access -> Subagent MCPs -> ?
next step: interactive (sub)agents that can communicate with the main agent. Seems like the perfect trade-off between context and composition/responsibility.
Yesterday I shipped "Agent Mode" in Sentry's MCP, which exposes the MCP's toolset as a single subagent.
I wrote up a bit more about this and the initial results (13,600 tokens -> 720 tokens, 2x response time, diff ergonomics).
cra.mr/subagents-with-mcp
1
161
asking LLMs to optimize prompts is like asking girls for dating advice... it just isn't good.
1
1
36
Would be amazing if claude code had a way to (dynamically) enable MCP usage to avoid cluttering the context.
Even better - make MCP tools available only to specific subagents, so work can truly be delegated.
@_catwu
3
69
Replying to @nickwm
@nickwm @cursor_ai please offer a larger subscription tier.
E.g. 2000-5000 premium tokens. Usage based pricing is not an option for a lot of workplaces.
26
Litmus test for AI agent frameworks: Does it provide proper interfaces for all major languages - if you can build great agents, you can surely automate your own integrations?
1
92
🚀 Just launched our biggest project at @passionfrootme – AI-powered search to find the perfect creators for your next B2B social media campaign! 🔍
Building a great search experience is hard.
Here’s what we learned 👇
An unified EU-wide company structure with with standardized investments, equity plans and employment sounds pretty nice eu-inc.org 🇪🇺
just found out you can do partial accepts in cursor. Super nice 🤯
just enable the setting and check the shortcut (⌘+→) to avoid those hallucinations :P
1
5
117
why is @OpenAI GPT4 still incapable of understanding its own API reference? Embarrassing
1
109
In case you didn't know how embeddings work yet, here is a simple infographic from #DALLE3, explaining each step😉
1
142







