Mukil Loganathan

112 posts

Mukil Loganathan banner
Mukil Loganathan

Mukil Loganathan

@MukilLoganathan

building @LangChainAI

Katılım Haziran 2016
201 Takip Edilen317 Takipçiler
Mukil Loganathan retweetledi
Viv
Viv@Vtrivedy10·
the real future of the very best vertical products is Model/Harness Choice + Openness easy to deploy infra is nice (great release from Ant) but it’s not the lever that matters the most at all to build the best agent when building a vertical agent you basically want to pull on every possible product lever to reach the Pareto frontier of Perf/Cost/Latency for your agent I want: - No model lock in, Claude is great but I need to access the best model for my task. If my competitor can have the same arch but access smarter, cheaper, models that I can’t, that’s bad. And the model wars aren’t stopping anytime soon. - Own my data and memory! Production data/evals are the currency of agent self-improvement. I want to be able to move/edit them somewhere else if that will make my agents better. - Finetune open, faster, specialized models on my data. I don’t want frontier intelligence on every task…I just want it on my task for my customers! The on-ramp to that is using your collected data to fine-tune a model for this. You might have to do this multiple times but it will still be cheaper than always routing to a frontier system we’re benefiting from an incredible intelligence battle between the labs and with open models open everything lets you choose when you want to pick a closed model, selfishly optimize for your product and promiscuously try every other choice :)
Jon Lai@Tocelot

a lot of talk on how 1000 startups just died due to Claude managed agents. I think that’s overblown - the truth is the moat for agentic products has been shifting from infra engineering to domain expertise + data for a while, managed agents is GOOD news if you’re a domain expert turned founder Previously if you were say a CPA building an agent - you had to wrestle with a lot of infra complexity just to get things working. Sandboxes for execution, state / session management, error handling etc Now with Claude handling the plumbing, you can focus on domain specific value creation, for example - tax logic / insights from the 1000s of returns you’ve processed as proprietary data for your agents to pattern match against - tribal knowledge on tax optimization strategies, state-specific quirks, weird behaviors that increase audit risk etc - things that aren’t found in generalized LLMs today - domain specific integrations into quickbooks for accounting, plaid for banking, avalara for tax reconciliation etc At end of day, for startups building vertical AI agents for “X” - what you really want to be selling is not the agent scaffolding itself but the codified expertise of a top practitioner of “X” - the judgement and outcomes of an expert doctor, accountant, lawyer, etc, encoded into software

English
3
9
46
6K
Mukil Loganathan retweetledi
Julia Schottenstein
Julia Schottenstein@j_schottenstein·
With the guy who needs constant supervision @LangChain’s leading model, Ezra
Julia Schottenstein tweet media
English
3
5
60
8.4K
Mukil Loganathan
Mukil Loganathan@MukilLoganathan·
@MeekMill this is the only right choice
Viv@Vtrivedy10

sometimes gotta just shoot your shot sometimes so @MeekMill whatever you need to keep cooking with AI, down to be your guy 🙏 - Temple grad (undergrad, masters, PhD in AI) 🍒 go owls baby 🦉 - been livin in Philly last 10 years, local (just over the bridge in Jersey don’t roast me) - will pull up anywhere you need to cook up AI stuff - work at LangChain, one of the companies that helped kickstart this AI revolution a few years ago, tryna actually make it way easier for everyone to use AI, not just the computer science bouls - go birds 🦅 (dh) - can spit Dreams & Nightmares if needed even if Meek doesn’t see this, it’s wild he cookin with AI 😅

English
1
0
6
198
Mukil Loganathan retweetledi
Viv
Viv@Vtrivedy10·
sometimes gotta just shoot your shot sometimes so @MeekMill whatever you need to keep cooking with AI, down to be your guy 🙏 - Temple grad (undergrad, masters, PhD in AI) 🍒 go owls baby 🦉 - been livin in Philly last 10 years, local (just over the bridge in Jersey don’t roast me) - will pull up anywhere you need to cook up AI stuff - work at LangChain, one of the companies that helped kickstart this AI revolution a few years ago, tryna actually make it way easier for everyone to use AI, not just the computer science bouls - go birds 🦅 (dh) - can spit Dreams & Nightmares if needed even if Meek doesn’t see this, it’s wild he cookin with AI 😅
MeekMill@MeekMill

Claude is helping me organize my whole music career and other businesses in days ... and it's moving my business forward at a high rate! Some tech youngbull I met on LinkedIn gave me a incredible template! Who else can help me with Claude

English
58
50
1.1K
340.1K
Mukil Loganathan retweetledi
Jesse Ellingworth
Jesse Ellingworth@JLEllingworth·
Built a v1 accounts payable agent with LangSmith Fleet today. Entirely in english, zero code. The agent reads my inbox for invoices, maintains a trusted vendor list, catches duplicate invoices, and forwards legit invoices to Ramp for further processing & approval. DMs me on Slack if anything needs a second look. Still working out the kinks, but should save our AP team at least an hour a day because we get a ton of invoice volume. We're just scratching the surface with agent use cases beyond coding.
LangChain@LangChain

Introducing LangSmith Fleet. Agents for every team. → Build agents with natural language → Share and control who can edit, run, or clone each agent → Manage authentication with agent identity → Approve actions with human-in-the-loop → Track and audit actions with tracing in LangSmith Observability Try Fleet: smith.langchain.com/agents?skipOnb…

English
3
15
35
7.2K
Mukil Loganathan retweetledi
Harrison Chase
Harrison Chase@hwchase17·
Open Models, Open Runtime, Open Harness - Building your own AI agent with LangChain and Nvidia Claude Code, OpenClaw, Manus and other agents all use the same architecture under the hood. They consist of a model, a runtime (environment), and a harness. In this video, we show how to create a completely open version of this: Open Models: Nemotron 3 Super Open Runtime: Nvidia's new OpenShell Open Harness: DeepAgents Video: youtu.be/BEYEWw1Mkmw Links: OpenShell DeepAgent: github.com/langchain-ai/o… Deep Agents: github.com/langchain-ai/d… OpenShell: github.com/NVIDIA/OpenShe…
YouTube video
YouTube
Harrison Chase tweet media
English
36
76
363
32.1K
Mukil Loganathan retweetledi
Mukil Loganathan retweetledi
Mukil Loganathan
Mukil Loganathan@MukilLoganathan·
Really excited to finally launch this in private preview. We have a lot more planned re memory, security, and monitoring to provide the most secure ephemeral environments for your agents. If you are interested would love to chat! We will be letting people off the waitlist incrementally but DM me if you want quicker access!
LangChain@LangChain

🚀 Today we're launching LangSmith Sandboxes Agents get a lot more useful when they can run code: analyze data, call APIs, build entire applications. Sandboxes give them a safe place to do it with ephemeral, locked-down environments you control. Now in Private Preview. Learn more: blog.langchain.com/introducing-la… Join the waitlist: langchain.com/langsmith-sand…

English
2
3
11
2.7K
Mukil Loganathan retweetledi
LangChain
LangChain@LangChain·
Deploy LangGraph agents using the LangGraph CLI You can now deploy LangGraph agents to production straight from your terminal using the LangGraph CLI! 🛠️ langgraph new → scaffold from a template 🧪 langgraph dev → test locally in Studio 🚀 langgraph deploy → deploy your agent on LangSmith 📋 langgraph deploy logs/list/delete → manage everything after directly from your terminal Blog: blog.langchain.com/introducing-de… Watch the full walkthrough: youtu.be/hcWHufkzicc
YouTube video
YouTube
LangChain tweet media
English
10
23
125
18.1K
Mukil Loganathan retweetledi
Viv
Viv@Vtrivedy10·
Harness Design Notes: Decoupling Agent Storage from Agent Compute TLDR: You can give each Agent/Subagent dedicated compute while sharing storage (repo/filesystem) to self-organize work between them. Shared Compute can be a bottleneck especially with long running code execution. Started writing up some harness design patterns over a very long flight this weekend, might make this a series if there's interest! We're on the edge of using a massive amount of compute to orchestrate agents across long horizon work Ex: for Agent Teams, an orchestrator organizes potentially many agents that fan out and do work on a project (like a large repo) For anyone who runs many agents locally, you see your CPU usage skyrocket for even moderate runs with code exec But Sandboxes to the rescue :) There's a nice pattern of shared filesystems via Volumes that all agents access while getting their own sandbox environment. The coordination happens via writing to the write place in the filesystem. And using git makes it so you can track and roll back changes over time good Harness Engineering on self-organizing agents via filesystems requires thinking about infra too. Many patterns work but you have to measure them for your work! Harness Engineering is Systems Engineering
Viv tweet media
English
21
32
377
19.9K
Mukil Loganathan retweetledi
Harrison Chase
Harrison Chase@hwchase17·
we're building ai into langsmith not just to be a generic assistant, but to actually help debug agents 🧵here's a real example where it helped me over the weekend: context: I'm building an agent on deepagents (github.com/langchain-ai/d…). It has a bunch of tools for interacting with files issue: I noticed thanks to langsmith monitoring (docs.langchain.com/langsmith/dash…) that ~1% of calls to `ls` were failing. sidenote - this is value of ai native monitoring, we automatically tracking failing tool calls. I clicked into an example run and saw that the model was generating the wrong parameter to `ls` - it was passing `file_path` not `path` at this point, i knew what the issue was, but had no idea WHY it was occurring. the trace here was very long and the prompt was long as well. i suspected that there was something wrong in the prompt - maybe a bad example? i asked polly (docs.langchain.com/langsmith/polly) our in app assistant to help me debug. she investigated, and found that other file tools in deepagents use `file_path`, and `ls` is the only one that uses `path`. see screenshot below I don't know how long it would have taken me to figure this out otherwise everyone is adding assistants into app for basic question/answering. imo really valuable assistants go beyond that - they are purposefully placed in situations where they can augment human intelligence nicely. in this case - reading long traces and prompts is something llms are great at!
Harrison Chase tweet media
English
12
7
66
8.5K
Mukil Loganathan retweetledi
Nuno Campos
Nuno Campos@nfcampos·
I wanted to share what we learned over the past few months, building agents 🧵
English
12
22
168
32.2K