Starnus helps founders and B2B teams run outbound with simple prompts. Define your ICP, find lookalike prospects, enrich with business + contact data, generate personalized outreach, and track replies, end to end in one platform.
I posted a random thread on X about the cost of living in the Netherlands. Nothing about what we're building. Just genuine thoughts about life in the Netherlands.
It hit 1M+ impressions. And here's the weird part we got a ton of signups and paid users for Starnus from it. Without ever mentioning the product.
Meanwhile, my "here's what Starnus does" posts? Way less engagement.
This genuinely messed with my head. I'm sharing the actual X post below
@osakasaul Very good point, we should indeed add an ICP builder
Actually good thing about Starnus is that it's like Claude, but for Sales and professional business work, so adding ICP builder means embedding such skills in the main AI brain (supervisor), since we already have 25+ specialized agents which cover web search, scraping and deep research, so we can do a comprehensive research, alognside a set of instructions that can help the system clearly define the ICP for the user.
Report
@osakasaul@khashayar_mansourizadeh1 ICP builders are where these platforms click or fall apart. If the matching logic can't handle vague descriptions against that profile database, it's just filters wrapped in a chat UI. Getting disambiguation right early is what would separate Starnus.
@mykyta_semenov_ very good point. we use "Unipile" as the LinkedIn automation engine/service, and over 1000s of reachouts we have had, no issues have happened. Unipile is one of the best solutions on the market, trusted for years by the market.
We always choose the best solution providers to integrate with Starnus.
Big advantage is that when you use Starnus, even the "Free Trial" or $24/mo plan, have full access to all premium databases, automation integrations, etc. which collectively can cost a single user $400+/mo if they should purchase them separately, plus the huge manual work overhead they would have.
Report
Does Starnus help with follow-ups automatically, or is it more about drafting and tracking?
@ititov_agency Absolutely, it's a completely intelligent entity with full capabilities of LinkedIn and Email automation, so it can definitely do follow ups and much more
The pivot from modular robotics to AI agent orchestration is a really compelling backstory. It takes guts to recognize when the market isn't moving fast enough for your hardware vision and shift entirely to software. I'm curious about how the agent marketplace works in practice, when a user submits a prompt for a sales task, how does Starnus decide which agents to orchestrate? Is there a ranking/reputation system for the agents, or is it more deterministic based on task type? Also, what's been the biggest surprise since shifting from hardware to the AI suite in terms of what customers actually want vs. what you expected?
@diegodau very good point. Here is the layered architecture:
1. Central brain and supervisor 2. Hyper professional skill sets for the supervisor 3. Agents registry
Supervisor decides which agents to use based on a combination of information, the registry data, what the skill set says, and the generic instructions it has. Registry contains the reputation and technical details.
The 25+ agents supervised by a central coordinator is interesting.
From an architecture standpoint, how do you manage latency when multiple agents run within a single campaign flow? Are they fully parallelized and event-driven, or is there queuing/step orchestration to prevent cumulative delays?
Curious how you balance agent specialization with responsiveness at scale.
We have built our own agent orchestration engine, protocol (similar to MCP, but fixed some scaling issues it had) and hub/registry, which allows us to run 10s of sequential batches, each containing 10s of agents which run in parallel. This means, in a single execution plan, we can run 100s of agents, make sure context is managed, flown through and orchestrated.
When we run multiple agents in 1 batch, they all run in parallel, and the results of the batch are passed to the next batch. The batch will wait, until the last agent has done the work, and then passes the context (which is not the fastest, but guarantees full dependency coverage).
Architecture is completely event-based.
Curious how you balance agent specialization with responsiveness at scale. --> I didn't exactly understand which aspect do you want to be compared? What do you exactly mean by responsiveness?
Report
@khashayar_mansourizadeh1 Thanks for the detailed breakdown that makes the batching model much clearer.
By “responsiveness,” I was mostly thinking from an end-user perspective.
If a single execution plan can involve 100s of agents across sequential batches, how do you ensure the user doesn’t experience noticeable delays when launching or monitoring a campaign?
Do you stream intermediate states/events back to the UI as batches complete, or does the user only see the result once the full orchestration finishes?
@rachid_jeffali Good point, it's rare that a single request would require 100s of agents, because agents themselves are layered, nested and some of them have 10-15 different modes of operation (so some agents are also a collection of sub-agents). I have executed (personally) single requests that required ~50 agents, but they were mostly ~5 agents executed for 10 different use cases.
We have live stream of events captured on UI, so users get updates every 5-15 seconds about the status of execution of each agent. This makes users more engaged and up to date about the progress.
The main supervisor is also a beast, that can execute batches, get back results, keep user posted, even update them by email, and move on. Set goals for future and keep working for days/weeks.
So there are many layers that try to keep the user engaged and updated, while the work is being done.
Report
Congrats on the launch! Moving from ICP to booked meetings in one flow is powerful. How does Starnus validate that the matched prospects truly fit the ICP beyond surface-level filters, especially in niche B2B markets where small mismatches can hurt conversion rates?
@vik_sh Starnus is integrated with 15+ databases, and powered by Claude on top of all, so when we find a lead, we dive into all details about them, even can analyze their websites and online presence, and decide from 0-100 how much they match the ICP we agreed on initially
You'll get an artifact database/list with all detailed info and analysis and scoring about each lead
RiteKit Company Logo API
Add an ICP builder wizard and you get the startup heads who don't have a degree in marketing.
What's nuts is that there's Gro launching today with a nearly identical value proposition.
I'm building too, and not really interested in building "yet another..." it seems that many are.
Starnus
@osakasaul Very good point, we should indeed add an ICP builder
Actually good thing about Starnus is that it's like Claude, but for Sales and professional business work, so adding ICP builder means embedding such skills in the main AI brain (supervisor), since we already have 25+ specialized agents which cover web search, scraping and deep research, so we can do a comprehensive research, alognside a set of instructions that can help the system clearly define the ICP for the user.
@osakasaul @khashayar_mansourizadeh1 ICP builders are where these platforms click or fall apart. If the matching logic can't handle vague descriptions against that profile database, it's just filters wrapped in a chat UI. Getting disambiguation right early is what would separate Starnus.
Starnus
@osakasaul @piroune_balachandran good point, indeed
With services like this, my main concern is always LinkedIn: what about bans for automation?)
For email, we use automated sending, but for LinkedIn we don’t — everything is done manually, because there’s a risk of losing the account.
Starnus
@mykyta_semenov_ very good point. we use "Unipile" as the LinkedIn automation engine/service, and over 1000s of reachouts we have had, no issues have happened. Unipile is one of the best solutions on the market, trusted for years by the market.
We always choose the best solution providers to integrate with Starnus.
Big advantage is that when you use Starnus, even the "Free Trial" or $24/mo plan, have full access to all premium databases, automation integrations, etc. which collectively can cost a single user $400+/mo if they should purchase them separately, plus the huge manual work overhead they would have.
Does Starnus help with follow-ups automatically, or is it more about drafting and tracking?
Starnus
@ititov_agency Absolutely, it's a completely intelligent entity with full capabilities of LinkedIn and Email automation, so it can definitely do follow ups and much more
ChatPal
Wow, seems powerful - you can then send linkedin messages and emails directly from platform?
Starnus
@daniele_packard yes exactly
Wordwand
The pivot from modular robotics to AI agent orchestration is a really compelling backstory. It takes guts to recognize when the market isn't moving fast enough for your hardware vision and shift entirely to software. I'm curious about how the agent marketplace works in practice, when a user submits a prompt for a sales task, how does Starnus decide which agents to orchestrate? Is there a ranking/reputation system for the agents, or is it more deterministic based on task type? Also, what's been the biggest surprise since shifting from hardware to the AI suite in terms of what customers actually want vs. what you expected?
Starnus
@diegodau very good point. Here is the layered architecture:
1. Central brain and supervisor
2. Hyper professional skill sets for the supervisor
3. Agents registry
Supervisor decides which agents to use based on a combination of information, the registry data, what the skill set says, and the generic instructions it has. Registry contains the reputation and technical details.
Wordwand
@khashayar_mansourizadeh1 I see, interesting! Thank you for answering me! Good luck
Starnus
@diegodau my pleasure, thanks!
The 25+ agents supervised by a central coordinator is interesting.
From an architecture standpoint, how do you manage latency when multiple agents run within a single campaign flow?
Are they fully parallelized and event-driven, or is there queuing/step orchestration to prevent cumulative delays?
Curious how you balance agent specialization with responsiveness at scale.
Starnus
@rachid_jeffali Good technical points!
We have built our own agent orchestration engine, protocol (similar to MCP, but fixed some scaling issues it had) and hub/registry, which allows us to run 10s of sequential batches, each containing 10s of agents which run in parallel. This means, in a single execution plan, we can run 100s of agents, make sure context is managed, flown through and orchestrated.
When we run multiple agents in 1 batch, they all run in parallel, and the results of the batch are passed to the next batch. The batch will wait, until the last agent has done the work, and then passes the context (which is not the fastest, but guarantees full dependency coverage).
Architecture is completely event-based.
Curious how you balance agent specialization with responsiveness at scale. --> I didn't exactly understand which aspect do you want to be compared? What do you exactly mean by responsiveness?
@khashayar_mansourizadeh1 Thanks for the detailed breakdown that makes the batching model much clearer.
By “responsiveness,” I was mostly thinking from an end-user perspective.
If a single execution plan can involve 100s of agents across sequential batches, how do you ensure the user doesn’t experience noticeable delays when launching or monitoring a campaign?
Do you stream intermediate states/events back to the UI as batches complete, or does the user only see the result once the full orchestration finishes?
Really interesting architecture approach.
Starnus
@rachid_jeffali Good point, it's rare that a single request would require 100s of agents, because agents themselves are layered, nested and some of them have 10-15 different modes of operation (so some agents are also a collection of sub-agents). I have executed (personally) single requests that required ~50 agents, but they were mostly ~5 agents executed for 10 different use cases.
We have live stream of events captured on UI, so users get updates every 5-15 seconds about the status of execution of each agent. This makes users more engaged and up to date about the progress.
The main supervisor is also a beast, that can execute batches, get back results, keep user posted, even update them by email, and move on. Set goals for future and keep working for days/weeks.
So there are many layers that try to keep the user engaged and updated, while the work is being done.
Congrats on the launch! Moving from ICP to booked meetings in one flow is powerful. How does Starnus validate that the matched prospects truly fit the ICP beyond surface-level filters, especially in niche B2B markets where small mismatches can hurt conversion rates?
Starnus
@vik_sh Starnus is integrated with 15+ databases, and powered by Claude on top of all, so when we find a lead, we dive into all details about them, even can analyze their websites and online presence, and decide from 0-100 how much they match the ICP we agreed on initially
You'll get an artifact database/list with all detailed info and analysis and scoring about each lead