←
Caspar411 posts
Caspar avatar✓

Caspar ✓

@caspar_br

Building Managed Agents at @LangChain 🇦🇺

◷ Joined June 2016▧ 48 Photos and videos
Caspar avatar
Caspar@caspar_br·20h

Running a workshop with @ishaan_lal1 today on Managed Deep Agents, come along to see what the future of agents looks like !

Caspar avatar
Caspar@caspar_br·Nov 16

We increasingly delegate decision making as models improve. In coding, we went from asking LLMs for: - a line of code ↓ a function ↓ a module ↓ a complete feature with /goal Next step is making this work in less verifiable domains. It will be real interesting to see the interfaces people build to enable feedback loops there

Caspar avatar
Caspar@caspar_br·Nov 16

Humans fall for scams. Agents do too. So how do you keep your data safe when your agent browses the web? In Managed Deep Agents you can now give the browser an allow list of hosts. If a page tells the agent to go anywhere else, the built-in proxy refuses the request Stay safe!!

Post attachment
Caspar avatar
Caspar@caspar_br·Nov 14

Cool lesson in agent design from Actively. How you structure your agents shapes how much they improve over time They shifted from one agent per customer to one per class of work, so lessons carry across customers and fewer corrections need repeating

Post attachment
Liam Bush avatar
⟳ Caspar retweeted
Liam Bush@LiamBush20·Nov 13

Series C baby! What a ride. From 60-400+ people I’ve been able to work with the most talented and tenacious individuals I’ve ever met. I love our odds. If you’re a lot more than a little insane we want you here!

Post attachmentPost attachment
1:57
Christian Bromann avatar
⟳ Caspar retweeted
Christian Bromann@bromann·Nov 13

“Still so much left to build” is the most accurate line here 🚀 Every week we ship to Managed Deep Agents, the list of what to build next gets longer, not shorter. 7,000+ companies on LangSmith will do that. Next chapter starts now. Grateful to build it with this team @caspar_br @sydneyrunkle @hwchase17 @huntlovell @ndrezn @VictorMoreira16 @masondrxy and more 😊

1:57
Caspar avatar
Caspar@caspar_br·Nov 13

When I joined LangChain we were just 80 people. Today we raised $300M at a $3B valuation My onboarding mentor was @sydneyrunkle, our open source titan. She taught me to care deeply about my craft. LangChain is many technologies built on top of each other, and Sydney was the first of many people here who hold every layer of that stack to a high bar Since then I've worked on our open source frameworks, then LangSmith Deployment, then Fleet, and now Managed Agents, where I now lead a team. The people here taught me that I don't have to stay in my lane. If something needs fixing, I can help fix it, wherever it lives in the stack We're now 400+ people with 7,000+ customers, and our new office is filling up fast! Thank you to @hwchase17 and @VictorMoreira16 for the drive to ship and to put customers first, to @ankush_gola11 for the confidence to tackle monumental projects, to @sydneyrunkle for the mentorship, and to my team for everything we build together This raise is about helping every company own its intelligence. It is a privilege to work towards it with some of the most talented and thoughtful people I know

Post attachment
1:57
Caspar avatar
Caspar@caspar_br·Nov 6

Managed Deep Agents can now be steered mid-turn. If you send a message while the agent is working, it reads it at its next model call instead of waiting for the turn to finish. Small change that makes agents much, much nicer to work with

Post attachment
LangChain avatar
⟳ Caspar retweeted
LangChain@LangChain·Nov 5

LangChain? LangGraph? Deep Agents? The full ecosystem explained in 3 minutes.

3:03
Jake Broekhuizen avatar
⟳ Caspar retweeted
Jake Broekhuizen@jakebroekhuizen·Nov 5

Some great mental models on writing high-quality evals by @austinberke and the harmonic team. A pleasure to work with them and some especially useful breakdowns of: - the tradeoffs between eval dimensions rather than optimizing for one score - evaluating verifiable correctness rather than prescribing a single "right" answer It also nails the importance of actually getting in and looking at the data to reward responses that build 'trust' - a quality that is often overlooked if the verification is purely left to a language model.