We want to hear from you (good and bad)
Quick context for those new here: Agently is a company brain your agents actually run on. You connect your tools, it builds one shared memory of how your company works, and the agents do the work on top of it. Not a chatbot. Something that compounds.
We're early. Two private cohort betas so far, and we're opened the public beta with our PH launch. The people who matter most to us right now are the ones already inside it.
So if you've tried Agently, come say hi. We wanna know:
What's the first thing you handed off to it?
What surprised you, good or bad?
What's still missing that would make you use it every day?
No polished answers needed. The rough stuff is the useful stuff, and we read all of it.
And if you're just finding us today: ask us anything in here. Omar and Ahmad (CEO/CTO) are both in the thread. 🚀
Replies
Haven’t tried it yet, but this thread is definitely making me want to. That line about it being not a chatbot but something that compounds” really sticks because most tools feel like they reset every time you open them. Would be interesting to see what it feels like after a few weeks of it actually learning how a company runs.
Agently
@tabitha_robyn Faster than that, actually. You connect your accounts and it backfills, your history is in the brain from day one, not built up over weeks. From there it scales as you do: the more your company runs through it, the more the brain holds, and your agents draw on all of it to actually do the work. It's not waiting to learn you, it starts with your context and compounds from there.
@omarships Really appreciate you explaining that so clearly. The compounding part makes a lot more sense now, it’s not catching up to your company, it already knows how it works.
the "shared memory that compounds" pitch is the interesting risk too though - if one agent writes something slightly wrong into that shared brain, does it just quietly become the assumption every other agent builds on top of? curious if there's any kind of confidence/source tracking on facts in that memory, or if it's treated as equally trustworthy regardless of which agent or tool contributed it
@omri_ben_shoham1Â Good question. In my own setup the memory is just plain text files on my Mac, so when the AI saves something wrong I can open the file and fix it like any other note. Once the memory is a box you can't open, a wrong fact just sits in there and everything builds on top of it. Curious how you guys handle corrections.
Agently
@omri_ben_shoham1 @mrnabeel Agreed on the instinct, if you can't fix the memory, a wrong fact just compounds. Two things handle it. First, it mostly doesn't get there: the brain runs continuous invalidation, so as new correct data flows in it supersedes stale facts on its own, a wrong one gets outranked instead of sitting there gathering weight. And when you do want to step in, you tell Jarvis it's wrong and the correction takes over from there, no digging through files. You don't hand-edit the store, but nothing's buried either: every fact carries its source and its history, so you can see what it believed and when. Structured beats a text file here, a correction doesn't just fix one note, it invalidates everything downstream that depended on it.
Agently
@omri_ben_shoham1Â No, it doesn't just silently become gospel. Every fact carries provenance, which agent or tool or human wrote it, and when. So facts aren't a flat pool of equally-trusted statements, the source travels with the data and is part of how conflicts get resolved. Two things stop the quiet-corruption problem you're describing. One: a new write doesn't overwrite, it competes. If an agent writes something that contradicts existing context, that's a flagged conflict, not a silent replacement, and provenance weighs into which one holds. Two: an agent can't write a consequential action into reality unattended anyway, that passes an approval gate. So a wrong fact shows up as a contradiction with a traceable source, not as a new assumption everyone silently inherits.
Haven't tried Agently yet either, but the question I'd want answered before handing anything off is about the shared memory itself - when it's wrong (stale info, or it picked up a bad pattern from how someone did something once), how do you go in and correct it? Is that a manual edit to the memory, or does it just get overwritten next time the correct process runs?
Agently
@galdayan Both, and you don't have to wait for it to fix itself. Two ways. One: the next correct run supersedes the old fact automatically. The memory's a temporal graph, so when new info contradicts something, it invalidates the stale version and marks it as no-longer-true rather than silently keeping both. Two: you correct it directly, tell Jarvis "that's wrong, here's the actual process," and your correction lands as a fact that outranks the bad one and invalidates it on the spot. Nothing gets hard-deleted either way, the old version stays in history marked stale, so you can see what it believed and when. And a bad one-off pattern doesn't calcify: it's one data point competing against everything else, not a rule until you make it one.
Every task has a human-in-the-loop feature, unless it is already set as recurring by the user or if it is a skill recognized by a pattern by the brain (also approved by the user).
this actually answers it, thanks - I was picturing a single mutable memory getting silently overwritten, not a temporal graph where the old fact stays visible as stale history rather than disappearing. the human-in-the-loop carve-out for anything non-recurring is the part I'd want to stress-test most in practice, curious how tight that gate feels once you're a few weeks into real usage rather than a demo.
@Omar Ghandour that provenance-plus-competing-write model makes sense, and the approval gate for consequential actions is the part that actually closes the loop for me. one follow-up: when two facts do conflict and both get flagged rather than silently resolved, who ends up looking at that queue in practice, is it whoever's using the agent at that moment, or does it route to someone specific on the team?
Agently
@omri_ben_shoham1Â Neither, actually, and that's kind of the point. A conflict doesn't fire off a ticket to someone. Both facts sit as competing edges and wait, and most resolve on their own: new data flows in and confirms one, so the other quietly falls away. No one has to sit in a queue. When it does need a human, it doesn't land as a "you have 1 conflict to resolve" alert. Jarvis just brings it up in conversation, as an elaboration rather than an error, "quick check, is X still the case or did it move to Y?" So it feels like it's clarifying, not like it's handing you a defect to triage. The resolution still carries provenance, but you experience it as the agent getting sharper, not as queue duty.
@omar_ghandour "it's clarifying, not handing you a defect" is a really good way to put it, that framing alone probably prevents a lot of the alert fatigue other memory tools run into. does the tone of that clarifying question ever change based on how long the two facts have been sitting in conflict, or is it always the same casual check-in regardless of how stale things get?