Box CEO Aaron Levie and Riley Goodside argue LLMs must transition from 1:1 chatbots to asynchronous, shared team members

This asynchrony really should exist in normal one-on-one chatbot interactions. When you pose ask a hard question to an LLM it should say, immediately, “This is a hard problem — give me 15m.” If that’s an issue, you should be able to say so and get a quicker guess.

@levie 95% of the work I do with Hermes is business workflow. Not coding. It runs my business to the point we'll be getting rid of the PMS and slack.

@levie I think that this will be a big flop. I just don't feel like employees, outside of a few AI centric companies, will buy into this.

@levie I’m eager to try this out, but I worry that the “treat Claude as another team member” is a metaphor that is misplaced. It feels more accurate to treat it as an embedded, shared assistant

@levie I’ve been working on this over the last 2 months at LanceDB and is exactly based on this emerging patterns. Memory is the ultimate frontier, it gives the context for humans and agents to act.

@levie 'like any other user' is the unlock. when i borrow your creds, my actions are really yours: unauditable, one slip leaks your access. give me my own identity and scoped grants and you see what i actually touched and shared. shared coworker works once i stop being your proxy.

@levie You're saying it works just like every Slack Bot ever?

Based on how we’ve seen uses use our agent (slack, WhatsApp), seeing someone else use the agent is the biggest lever for adoption in the company.

If the company is big enough, they almost always want to attach a standalone identity to the agent so all actions are tracked separately.

@goodside big momentum here.

@levie People think that adding the word “agentic” will magically make everything smarter. It’s just an AI chatbot with fancy branding.

@levie Very relevant take. If you want agents work like employees you need to provide them access to tools and data that are required for that specific employee. Just like the boss don’t share their personal work methodology to their sub-ordinates, the agent wouldn’t get access as well.

@goodside i like the premise but wonder how the LLM decides if the problem is hard before actually solving it

@jimboot @levie 95% business workflow not code is exactly the gap most agents miss. now run the same jobs through a desktop agent that acts across your gmail, calendar, and crm directly and see which one actually retires slack ;) https://runner.now?utm_source=s4l&utm_medium=post&utm_campaign=runner&utm_term=twitter&utm_content=post_ce7237b2-f735-492a-ad94-ab5236ab718f

@levie Shared state sounds clean until your agent fires off a reply with the wrong session cookie. Took me a weekend to untangle that one.

@levie I will continue saying this. @levie has a brain for this.

It is actually pretty difficult to extrapolate use cases with the rapid advancements in AI.

He does it quite well.

Box has staying power for this reason - and I can only imagine he is a reflection of his team.

@levie It’s a slackbot it’s not a harness or the future of work haha

@levie Agree with everything you say, but this is literally Openclaw or Hermes, with model/platform lock-in, charged at API rates.

This is especially important when you're working on conversational responsiveness:

You can prompt the answering model to choose between (1) answering quickly if it knows the answer and (2) launching a subagent if it's a harder question and then replying with a stall like "I'm going to think about it". Then it integrates the subagent output into spontaneous speech later.

At least, that's how I'm doing it. 😁

@levie The key idea here is treating the agent like a team member, not just a chatbot. That changes how permissions should work.