Elvin
Proactive AI that finds and finishes work before you ask
480 followers
Proactive AI that finds and finishes work before you ask
480 followers
Stop acting as the routing layer between AI and your actual work. Elvin is proactive AI that finds coordination work across your tools, handles the messy multi-step parts, and asks before taking action. It turns scattered context from messages, meetings, docs, and task tools into ready-to-approve drafts, follow-ups, updates, and next steps.
The proactive angle is compelling. Most AI tools still wait for a prompt, which keeps humans as the workflow router. How does Elvin decide when to start work on its own versus when to pause and ask for confirmation? That boundary feels like the difference between an assistant and a real work agent.
Elvin
@rahulbhavsar The question about how proactive to be is a really great one, and it's one that we've shifted as we've built and tuned the product. Here's the model that we're working with now. For novel tasks that you give Elvin or Elvin finds, we tell Elvin that it can "read" as many tools as you've given it access to, then it prepares a plan for how it will do the work. Then the user needs to approve the plan or request a change. Elvin then executes the task. If it runs into an issue in the execution and needs to request additional permissions (reading is still "free"), then it needs to come back to the user for another approval.
We also allow you to set up recurring skills (for example, reviewing all your meeting notes at the end of the day and compiling action items). Once those are set up, they run with the pre-approved plans!
It's also something we may continue to revisit.
How do you think it should work?
@rahulbhavsar @ellie_powers "Reading is free, acting requires approval" feels like the right default. The boundary between assistant and agent is exactly there. If an AI starts taking actions before showing the plan, it's not really reducing the cost of work - it's just transferring the risk to the user.
@jared_salois Exactly “Reading is free, acting requires approval” is a strong default. I’d add that recurring tasks can be pre-approved only when the playbook is narrow and the audit trail is clear. Otherwise autonomy can look productive while quietly increasing risk.
@ellie_powers I like that model. My instinct is to separate work into four levels: read/understand, draft/plan, reversible actions, and irreversible or customer-facing actions. The first two should be free, reversible actions can be pre-approved inside a narrow playbook, and anything irreversible should need human approval plus an audit trail. That feels like the trust boundary between an assistant and an AI employee.
I like the idea that Elvin looks for the coordination work first instead of waiting for a perfect prompt. Emails, meetings, docs, and task tools are exactly where the messy “someone needs to follow up on this” work usually lives.
The approval step also feels important. For anything involving teammates, clients, or important decisions, I would want the agent to prepare the work, but not silently act without me.
Curious how Elvin decides what is worth surfacing. Does it learn from what users approve, ignore, or edit over time?
Elvin
@andrasczeizel Thank you for learning about Elvin! Yes, it does learn from what users approve and ignore - there is also a button on items that Elvin found where you can tell it that it identified an item you'd like it to skip in the future and it will learn from that. We're definitely tuning it over time so that it will get closer to perfect results for each user as soon as possible.
Lancepilot
Elvin
@odeth_negapatan1 Thank you! Yes, we are excited about a future where human minds and capacity are supported by AI, not overwhelmed by the noise it can generate. I would say that one of the most impactful surprises early on was on @workmonk 's account. He received an invitation to speak at a conference. Elvin offered to put together a speaker proposal. It looked in his personal context layer, talking about his specific technical experience, Granola meeting notes, Linear tickets, emails, and more, to come up with a meaningful and technically accurate speaker proposal. All he had to do was review the plan and say yes, and boom! A better speaker proposal than he thinks he would have generated on his own. Our minds were blown.
Proactive AI that surfaces things before you ask is such a harder problem than reactive AI — you need to deeply understand context, priorities and timing. Really curious how Elvin decides when to interrupt vs stay silent. Any plans for calendar/email integration? Congrats on the launch!
The Contribution License feels a bit worrying. I wouldn't use this on your work email unless you're looking for trouble.
It basically says that anything you submit/post can be used, copied, stored, sold, published, sublicensed, translated, and exploited for commercial or advertising purposes.
@Elvin Congrats on the launch. Proactive AI that finds coordination work before the user asks is a strong direction. One QA edge case I’d watch is approval visibility: what changed since the plan was created, which tools will be touched, and how users recover if the result is wrong. Want a few agent-approval QA scenarios?
Empromptu AI
Elvin
@shanealeven Thank you, Shanea! We really appreciate your support. :-)