Introducing Agently.dev: An Autonomous Brain Connecting Your Company’s Entire Tech Stack

·Jul 29, 2026 06:32 PM·

Hey Everyone, Omar here. Multi-exited founder, ex-investment banker and consultant in a previous life. Now building Agently.dev: an autonomous brain for your company. The short version: it connects your whole stack (email, CRM, docs, calendar) into one living graph that never forgets, and an orchestration layer dispatches agents that run work end to end. What comes out is real artifacts, not chat logs. 550 Teams already run on it, from solo founders through mid-market to our one enterprise client, and we launched on Product Hunt earlier this month and got #4. Happy to swap notes on agents, GTM, or anything early-stage, and always up for honest feedback. What are you all building? X: x.com/omarships LinkedIn: linkedin.com/in/omarghand

💙1

👋1

5 comments

· Sorted by Oldest

jonatan
·
·
welcome Omar. artifacts instead of chat logs is the right call, that's the line between something that demos well and something a team actually depends on. the part I'd expect to be hardest isn't the graph, it's deciding what's allowed to be written into it, because once agents can write freely it fills up with things that were true for one afternoon and the memory starts working against you. 550 teams and a top five launch is real traction, congrats.
Omar G.
·
·
what nice is that we abstract away as much as possible agents writing into the brain. The graph is based of relationships and not summaries or inference. When the data comes in we normalize, embed and validate, all without needing any agent yet. Also, through the use of the brain either through MCP (coming soon) or through the Work OS, we use RLHF as another arbitrator of knowledge
jonatan
·
·
relationships rather than summaries is the right call. the summary is where a lossy opinion sneaks in and then everything downstream inherits it without knowing it did. the one i'd think about next is time. relationships are facts with an expiry date. someone was the buyer, a tool was in the stack, a deal was open. validating on ingest tells you it was true when it arrived, not that it's still true, and a graph that never forgets is also a graph that never admits an edge went stale. so the useful bit isn't only what it knows, it's how old each edge is and whether the system will say so out loud. and i'd be careful leaning on rlhf as an arbitrator of knowledge. feedback tells you which answers people liked, which isn't the same as which ones were correct, and in gtm those two come apart most exactly where the stakes are highest. a confident wrong answer gets a thumbs up more reliably than an honest i don't know.
Omar G.
·
·
our relations are temporal, so defintiley already does what your inferring. As for RHLF very good point, although it's not an arbitrator of knowledge but "another" arbitrator of knowledge. Meaning dissonance will definitely be flagged
jonatan
·
·
temporal edges cover the half where the graph knows when something was true. the harder half is an edge that's old but never contradicted, because most systems read no contradiction as still true, and then a two year old relationship and a two day old one come back looking identical at query time. the age has to show up in the answer, not just sit in the record. and another arbitrator is the fairer framing, agreed. the thing dissonance catches is sources disagreeing. what it won't catch is quiet agreement that's wrong, which in gtm data is the more common one, everything lines up because everything was copied from the same original record.

jonatan
·
·
welcome Omar. artifacts instead of chat logs is the right call, that's the line between something that demos well and something a team actually depends on. the part I'd expect to be hardest isn't the graph, it's deciding what's allowed to be written into it, because once agents can write freely it fills up with things that were true for one afternoon and the memory starts working against you. 550 teams and a top five launch is real traction, congrats.
Omar G.
·
·
what nice is that we abstract away as much as possible agents writing into the brain. The graph is based of relationships and not summaries or inference. When the data comes in we normalize, embed and validate, all without needing any agent yet. Also, through the use of the brain either through MCP (coming soon) or through the Work OS, we use RLHF as another arbitrator of knowledge
jonatan
·
·
relationships rather than summaries is the right call. the summary is where a lossy opinion sneaks in and then everything downstream inherits it without knowing it did. the one i'd think about next is time. relationships are facts with an expiry date. someone was the buyer, a tool was in the stack, a deal was open. validating on ingest tells you it was true when it arrived, not that it's still true, and a graph that never forgets is also a graph that never admits an edge went stale. so the useful bit isn't only what it knows, it's how old each edge is and whether the system will say so out loud. and i'd be careful leaning on rlhf as an arbitrator of knowledge. feedback tells you which answers people liked, which isn't the same as which ones were correct, and in gtm those two come apart most exactly where the stakes are highest. a confident wrong answer gets a thumbs up more reliably than an honest i don't know.
Omar G.
·
·
our relations are temporal, so defintiley already does what your inferring. As for RHLF very good point, although it's not an arbitrator of knowledge but "another" arbitrator of knowledge. Meaning dissonance will definitely be flagged
jonatan
·
·
temporal edges cover the half where the graph knows when something was true. the harder half is an edge that's old but never contradicted, because most systems read no contradiction as still true, and then a two year old relationship and a two day old one come back looking identical at query time. the age has to show up in the answer, not just sit in the record. and another arbitrator is the fairer framing, agreed. the thing dissonance catches is sources disagreeing. what it won't catch is quiet agreement that's wrong, which in gtm data is the more common one, everything lines up because everything was copied from the same original record.

Introducing Agently.dev: An Autonomous Brain Connecting Your Company’s Entire Tech Stack | RevGenius