nathan drezner

PM @LangChain building Managed Deep Agents. I post demos, code, and what breaks. prev @plotly

nathan (at) drezner (dot) xyz

[nathan drezner @ndrezn]

Browser use w/ @LangChain + @typesafeai's Jev! Really fun to build.

... I found it's excellent at playing the Wikipedia Game. (But, it's also great at "folding laundry" type tasks, like finding cheap flights.)

[Jeff Weinstein @jeff_weinstein]

LangChain developers can now easily build agents that can pay for things with @link (which works across any checkout online or programmatic payments over @mpp).

See their in-depth guide and demo! =>

[nathan drezner @ndrezn]

source code (with a rehearsal mode, so you can easily take it for a spin without real purchases): https://t.co/CnMYGGpg4T

[nathan drezner @ndrezn]

we built Restock, an agent to automate ordering office supplies using @link by @stripe - pens, groceries, and more. it's open source, so you can fork it, modify it, and deploy it yourself with `mda deploy`

payments are handled via @mpp. Restock is a great example of a complex use case which handles user authentication, secure secrets storage, and human-in-the-loop payments approval.

source code 👇

[LangChain @LangChain]

You can now build an agent that:
🔎 Finds real products
🛒 Prepares purchases
💳 Pays
…with a person approving every order 👨‍💻

Ask in @slackhq, review the order, approve in @Stripe's Link agent wallet.

Built on MPP + Managed Deep Agents.

https://t.co/UY3PTfdWsX

[Cohere @cohere]

We’re announcing Cohere’s first office in Ottawa 🇨🇦

Canada is where Cohere started, and we’re proud to keep building here.

[nathan drezner @ndrezn]

Really sad to hear that Margaret Hamilton has passed away. Led an incredible life and was an inspiration to so, so, so many. From MIT: https://t.co/ndlwsqc5hp

[nathan drezner @ndrezn]

Here are the full details of the release and how you can get started: https://t.co/i9QYfzz9zX

[nathan drezner @ndrezn]

We just shipped 🦾 MDA v0.9 - it adds dynamic agent definitions 🧱, Slack reactions 👀, and...

It's now possible to build agents which can dynamically set schedules.

Build tools for your agents around the new schedules SDK to let anyone using their agent schedule updates, deep research tasks, and more.

Happy building 🏗️

[nathan drezner @ndrezn]

Really excited to try building against this protocol - there are a lot of interesting explorations in this space like web MCP and MPP.

but there's a lot of innovation that can happen in this space and PAP feels like a step in the right direction

[Sierra @SierraPlatform]

Today we’re announcing Personal Agent Protocol — an open standard @Meta and @SierraPlatform are developing along with industry partners at @Genesys, @instinct, @RocketOTD, @Shopify, @stripe, and @Walmart. It will help define how personal agents interact with businesses and is open for anyone to implement.

Read more: https://t.co/9cWCVlCv8H

[Caspar @caspar_br]

let me introduce you to the langchain stack

1. langgraph gives you this level of control
2. traces give you the complete run lineage
3. version your skills and instructions in context hub
4 & 5. deepagents has durable execution built in and manages the context window for you

[Caspar @caspar_br]

time to level up https://t.co/pnmD06la9i

[nathan drezner @ndrezn]

free web search! for your agents! backed by Parallel with ZDR. Take it for a spin today.

[LangChain @LangChain]

Add web search to your Managed Deep Agents for free.

🌐 Powered by @p0

💻 No separate vendor accounts, managing extra PI keys, or wiring search tools necessary

A quick demo from @ndrezn

[nathan drezner @ndrezn]

highly encourage reading the full paper, the best way to learn is to see what we've tried already in the past: https://t.co/Zv6FBEMSOR

[nathan drezner @ndrezn]

morning reading today is Lifestreams (1996) which Scott Jenson referenced in his excellent talk on innovating the desktop OS -

lots of interesting ideas in this paper.

- implicit organization: our interactions with the computer say more about intention than we might realize in terms of how information should be organized

- filesystem management as a first stem is a poor abstraction: we never write the name of the project on a piece of paper nor are we required to stuff it in a folder in real life, so why do we require this in our digital lives?

- document-first & compression-first design: our interactions with computers consist of the production of documents. let's optimize the computer to sort and find information out of this mind map

I think Obsidian might be the closest modern human-first implementation of a system like this but there are a ton of ideas in here I think are fascinating when thinking specifically about agentic memory and retrieval systems that are possible today with llms

[nathan drezner @ndrezn]

the proposal of "personal agents" as first-class citizen is likewise prescient - different agents designed for different functions.

in this paradigm a personal agent is defined as a specific function to avoid having to delegate common workflows to dedicated applications. similar perhaps to "skills" tools for llms today, or perhaps a complex agent in this paradigm might be akin to today's purpose-built agents (like what folks can build with deepagents)

[nathan drezner @ndrezn]

huge congrats to Polars team - looking forward to using v2!

[polars data @DataPolars]

We just released Polars 2.0.

It removed many of our legacy decisions makes the streaming engine our default and promotes SQL to a first class citizen within Polars.

It comes with initial out-of-core (spill to disk) support, a new Map data type and a lot of performance improvements. In fact, we think Polars is now one of the fastest analytical SQL engines on a single node. See benchmarks in the post: https://t.co/ATc6UzJqby

[Caspar @caspar_br]

User memory in Managed Deep Agents 0.8: your agent can remember the people it works with https://t.co/pnmD06la9i

[nathan drezner @ndrezn]

good take on HN about the value of durable harnesses re Pi Durable - very excited to try out the new Pi harness

[nathan drezner @ndrezn]

great piece from @sydneyrunkle on moving model routing upstream into the harness -

great way to think about managing routing more efficiently and building your harness to avoid context bloat.

[Sydney Runkle @sydneyrunkle]

https://t.co/sGbf2IE6q0

[nathan drezner @ndrezn]

@sivareddyg @MistralAI Congrats Siva!!! Excited to see what you work on