Most AI tools help you write. Ours helps you do research.
Today we launched the AI Co-Researcher at Zerve.
It's a multi-agent system that runs the research cycle end to end. Not a chat assistant sitting next to your notebook. A parent agent that plans the work and spins up sub-agents for systematic review, coding, analysis, replication, and drafting.
#AI #MultiAgent
Today we're launching the Agentic Report on @Zerve_AI
Analysis got faster. Communicating it didn't.
One prompt. The agent turns any analysis into a live report stakeholders can chat with.
Same analysis. CIO version. PM version. Risk version. Three reports, three audiences, one source of truth.
We added a Deploy tab to notebooks.
The agent reads your code, writes the app, gives you a URL. FastAPIs, Flask, Streamlit, Gradio, Dash — same project.
Hot reloads. Live previews. Streaming logs.
@Zerve_AI turns any analysis into a live app in seconds.
Full demo below ↓
Notebooks feel brittle because meaning moves, but the name doesn’t.
A small working set of variables evolves, gets reused ~3.4× more, lives ~10× longer, and is depended on far downstream.
Local change. Global dependency.
Is AI and vibe coding changing what actually gets shipped?
I looked at recent Product Hunt data to find out.
AI Marketing, Dev Tools, and Productivity now make up ~55–60% of AI launches. High volume. Heavy competition.
Some categories are accelerating fast. AI Analytics ~80% growth. AI Writing ~76%. AI Design ~60%.
But activity isn’t outcome.
AI products get ~60% more average votes, while the median stays the same.
The standout is AI Voice & Speech. Lowest noise. Highest winner rate (~23%).
If launches are an early signal, what does this say about 2026?
8/🧵 There are interesting exceptions.
For some companies, like Perplexity and Sierra, founder attention shows a positive relationship with traffic.
These cases look less like isolated spikes and more like coordinated narrative reinforcement across founder, product, and company.
That suggests a different dynamic:
attention not just as awareness, but as part of a broader go-to-market system.
1
1
156
7/🧵 One thing I tested explicitly: traffic (the general case).
Across founders, attention spikes do not reliably correlate with traffic or search interest.
In AI-first markets where SEO is weak early, founder posting does not substitute for search.
1
1
18
6/🧵 Another clean separation appears:
Authenticity drives engagement.
Authority drives attention.
They serve different purposes, and conflating them tends to hurt both.
1
1
13
3/🧵 Followers still matter, but mainly as a floor.
Across founders, follower count explains ~30% of expected median impressions.
It sets baseline visibility.
It does not explain who ends up at the top of the distribution.
1
1
23
1/🧵 Founder-led marketing follows a power-law.
Looking at founder posts across AI companies over the past 12 months:
~1% of posts account for ~49% of total impressions
• ~5% account for ~74%
• ~20% account for >90%
Winner-takes-most dynamics, similar to venture returns.
1
1
190
5/🧵 But this is conditional.
When the information doesn’t exist — private data, internal schemas, unfamiliar systems — there’s nothing to retrieve.
Other agents have to discover the structure through execution.
1
13
🧵/4 In some domains, the system prompt can be optimised so that prompts of very different fidelities produce the same behaviour.
This works because the model already holds the relevant knowledge.
The agent is instantiating, not discovering.
1
16
3/🧵 Embedding similarity between inputs didn’t predict behaviour.
What embedded closely was the plan, once a system prompt was applied.
Different semantic prompts.
Same plan.
1
16
2/ 🧵 What I expected:
Certain words would carry most of the information.
“Build a Spotify clone” should embed very differently from a vague prompt.
That wasn’t what happened.
1
35












