Not another AI bot, a real colleague to work alongside your team. Give your team superpowers or even run your company on autopilot.
Replies
Best
Congrats on the launch, Dan! 🚀 It’s awesome to see you guys pulling back the curtain on how Scarlett runs. The decision to skip the traditional Vector DB route and stick to an SQL-heavy setup for speed (similar to Hermes) is a really fascinating engineering choice.
Since you mentioned Vector DBs sometimes over-cluttered the context, how do you handle deep semantic search or long-term recall when a user asks about a conversation from months ago? Is it purely structured querying, or do you have a hybrid text-search layer built into the SQL database?
I asked Scarlett to explain it herself... here's what she said:
"Good question, this was one of the trickier calls. Short answer: it's a hybrid. We didn't drop semantic search, we just folded it into the SQL layer instead of running a separate vector DB.
Postgres full-text search handles most "find where we talked about X" recall. For fuzzier matches, we keep vector similarity as a column/index right next to the structured rows, so one query can filter on structure (who, when, which entity) and then rank by semantic closeness. That structured pre-filter is what kills the over-clustering problem: we prune the candidate set before semantic ranking, so we don't drag in the loosely-related noise a pure top-k vector lookup would.
For long-term recall we don't re-scan everything, raw messages get compiled into entity-scoped, source-backed memory, so old questions resolve against compiled facts first and drill into raw evidence only when someone needs the exact quote.
So: structured querying as the backbone, with lexical and vector layers riding inside the same engine. Speed and clean context were the whole point."
So we have a Karpathy style wiki for long term, but it's stored in the db just structurally.
Report
I like that you're positioning Scarlett as a co-worker rather than another AI assistant. That's a much higher bar, because teammates aren't judged by how much they can do—they're judged by whether people trust them with responsibility. If AI reaches that point, the relationship users have with these products changes completely.
@aryan787544 Absolutely. We work remote so I treat her pretty much the same as the humans on the team (though expect a little more). FWIW we tried doing multiple agents like one for engineering, one for marketing, etc. And I've seen some companies try 1 agent per person, but i think just a single Scarlett is simplest.
Report
@dansutera That's interesting because it suggests the product decision wasn't really about agent architecture—it was about reducing the user's management overhead. One capable co-worker gives people a single relationship to build trust with, whereas multiple specialists risk making users manage a team instead of getting work done. It'll be interesting to see whether that simplicity becomes a long-term advantage as agent capabilities improve.
Report
This is neat. Does it keep context across different Slack channels or only within threads?
@dhiraj_patel5 Hey yes, you get context from any channel you want. You can even give it access to your private chats if you so choose.
Report
The iMessage angle is underrated for solo operators. A lot of small businesses do not live in Slack, and forcing them into a new cockpit can kill adoption. The piece I would want very visible is the control surface: what Scarlett is allowed to do alone, what needs approval, and what happened while nobody was watching.
@krekeltronics Yeah, we found a lot of people that we talked to weren't on slack, especially solo & small companies. We do have a basic list of upcoming and completed tasks in the dash. But you can also ask Scarlett to make additional daily reports that give you the details you're looking for...
Report
@dansutera This one's personal. I've been duct-taping AI tools together at Ctruh to cover exactly this gap, and an AI teammate that actually holds context and owns the ops glue is the dream. Congrats team, taking it for a spin this week. 🔥
@sharun_kanan Hey thanks! Yeah we used to be on OpenClaw but it was too duct-tapey and had a mind of its own. We are running primarily on OpenAI's Agent harness which is a lot more reliable. Let me know how it is for you!
Report
Congratulations on the launch!
Scarlett sounds like a practical AI agent rather than just another chatbot. how does she decide when to act autonomously versus asking the user for approval, especially in business workflows?
the Autopilot capability sounds powerful. for founders who are just getting started, what are the first three workflows you would recommend automating with Scarlett to see value quickly?
@imogen_wallace Thanks Imogen! Autopilot is great yeah. Workflow-wise i like to get daily company report going. Daily customer triage. And automating whatever your main social is such as X.
Great product guys, love the fact it comes with 500+ connectors. Let's goo!
Report
Congratulations on the launch!
i really like the idea of using your API keys instead of requiring users to manage multiple subscriptions. how do you handle usage limits and cost optimization behind the scenes?
@james_carter35 Thanks james! It's tricky, depends on the platform. Optimally there is pay per use with auto recharge like XAI. Other ones sometimes need multiple accounts for cost optimization.
Replies
Congrats on the launch, Dan! 🚀 It’s awesome to see you guys pulling back the curtain on how Scarlett runs. The decision to skip the traditional Vector DB route and stick to an SQL-heavy setup for speed (similar to Hermes) is a really fascinating engineering choice.
Since you mentioned Vector DBs sometimes over-cluttered the context, how do you handle deep semantic search or long-term recall when a user asks about a conversation from months ago? Is it purely structured querying, or do you have a hybrid text-search layer built into the SQL database?
Scarlett.
@franz_briones Thanks Franz!
I asked Scarlett to explain it herself... here's what she said:
"Good question, this was one of the trickier calls. Short answer: it's a hybrid. We didn't drop semantic search, we just folded it into the SQL layer instead of running a separate vector DB.
Postgres full-text search handles most "find where we talked about X" recall. For fuzzier matches, we keep vector similarity as a column/index right next to the structured rows, so one query can filter on structure (who, when, which entity) and then rank by semantic closeness. That structured pre-filter is what kills the over-clustering problem: we prune the candidate set before semantic ranking, so we don't drag in the loosely-related noise a pure top-k vector lookup would.
For long-term recall we don't re-scan everything, raw messages get compiled into entity-scoped, source-backed memory, so old questions resolve against compiled facts first and drill into raw evidence only when someone needs the exact quote.
So: structured querying as the backbone, with lexical and vector layers riding inside the same engine. Speed and clean context were the whole point."
So we have a Karpathy style wiki for long term, but it's stored in the db just structurally.
I like that you're positioning Scarlett as a co-worker rather than another AI assistant. That's a much higher bar, because teammates aren't judged by how much they can do—they're judged by whether people trust them with responsibility. If AI reaches that point, the relationship users have with these products changes completely.
Scarlett.
@aryan787544 Absolutely. We work remote so I treat her pretty much the same as the humans on the team (though expect a little more). FWIW we tried doing multiple agents like one for engineering, one for marketing, etc. And I've seen some companies try 1 agent per person, but i think just a single Scarlett is simplest.
@dansutera That's interesting because it suggests the product decision wasn't really about agent architecture—it was about reducing the user's management overhead. One capable co-worker gives people a single relationship to build trust with, whereas multiple specialists risk making users manage a team instead of getting work done. It'll be interesting to see whether that simplicity becomes a long-term advantage as agent capabilities improve.
This is neat. Does it keep context across different Slack channels or only within threads?
Scarlett.
@dhiraj_patel5 Hey yes, you get context from any channel you want. You can even give it access to your private chats if you so choose.
The iMessage angle is underrated for solo operators. A lot of small businesses do not live in Slack, and forcing them into a new cockpit can kill adoption. The piece I would want very visible is the control surface: what Scarlett is allowed to do alone, what needs approval, and what happened while nobody was watching.
Scarlett.
@krekeltronics Yeah, we found a lot of people that we talked to weren't on slack, especially solo & small companies. We do have a basic list of upcoming and completed tasks in the dash. But you can also ask Scarlett to make additional daily reports that give you the details you're looking for...
@dansutera This one's personal. I've been duct-taping AI tools together at Ctruh to cover exactly this gap, and an AI teammate that actually holds context and owns the ops glue is the dream. Congrats team, taking it for a spin this week. 🔥
Scarlett.
@sharun_kanan Hey thanks! Yeah we used to be on OpenClaw but it was too duct-tapey and had a mind of its own. We are running primarily on OpenAI's Agent harness which is a lot more reliable. Let me know how it is for you!
Congratulations on the launch!
Scarlett sounds like a practical AI agent rather than just another chatbot. how does she decide when to act autonomously versus asking the user for approval, especially in business workflows?
Scarlett.
@avery_thompson2 thanks Avery!
The autopilot is autonomous, but you can turn on or off and change frequency. Autopilot will ask for help / approval if it thinks its necessary.
Normal scarlett mode (autopilot off) is always acting only when asked.
Dune
Congratulations on the Launch!!
Scarlett.
@dhanrajchoudhary Thanks Dhanraj!
the Autopilot capability sounds powerful. for founders who are just getting started, what are the first three workflows you would recommend automating with Scarlett to see value quickly?
Scarlett.
@imogen_wallace Thanks Imogen! Autopilot is great yeah. Workflow-wise i like to get daily company report going. Daily customer triage. And automating whatever your main social is such as X.
Microlaunch
Great product guys, love the fact it comes with 500+ connectors. Let's goo!
Congratulations on the launch!
i really like the idea of using your API keys instead of requiring users to manage multiple subscriptions. how do you handle usage limits and cost optimization behind the scenes?
Scarlett.
@james_carter35 Thanks james! It's tricky, depends on the platform. Optimally there is pay per use with auto recharge like XAI. Other ones sometimes need multiple accounts for cost optimization.