ทวีตที่ปักหมุด
Side by side example
Same model (claude-opus-4-6). Same task. Two different agent harnesses
@LangChain Deep Agents CLI: 9s
Claude Code: 16s
The harness IS the performance. 1.7× difference, zero model changes
Viv@Vtrivedy10
English
Git Maxd
3.8K posts
Git Maxd
@GitMaxd
Linux Dev | AI Rev Early Adopter | Next | Gen AI\UX | Deep Agent Fleets
Phoenix, AZ เข้าร่วม Temmuz 2023
477 กำลังติดตาม1.8K ผู้ติดตาม
@helloiamleonie Make agents send text/markdown headers first
Make websites deliver progressively disclosed markdown content in >500 tokens, like an agent can quickly and cheaply consume to pick the next discovery path
English
Noise over signal seems normal rn - the noise is the marketing.
Agents don’t care.
They want the data, they shouldn’t fetch HTML is their only surface - it should be the fallback surface in the event that a website does not provide a structured, progressively disclosed markdown response instantaneously
English
@GitMaxd @karpathy @ryancarson @RhysSullivan It does feel like there are so many feels ppl assume the agents looks at and follows, when in reality, we’re really just making files that we as humans read and understand, and agents ignore 😜
English
Thank you Sarah, my pleasure to come on the pod! And happy to do some more Q&A in the replies.
sarah guo@saranormous
Caught up with @karpathy for a new @NoPriorsPod: on the phase shift in engineering, AI psychosis, claws, AutoResearch, the opportunity for a SETI-at-Home like movement in AI, the model landscape, and second order effects 02:55 - What Capability Limits Remain? 06:15 - What Mastery of Coding Agents Looks Like 11:16 - Second Order Effects of Coding Agents 15:51 - Why AutoResearch 22:45 - Relevant Skills in the AI Era 28:25 - Model Speciation 32:30 - Collaboration Surfaces for Humans and AI 37:28 - Analysis of Jobs Market Data 48:25 - Open vs. Closed Source Models 53:51 - Autonomous Robotics and Atoms 1:00:59 - MicroGPT and Agentic Education 1:05:40 - End Thoughts
English
I think it seems kind of clear at this point
Token$ are almost irrelevant in this phase - everyone has limitless subs and the providers have no incentive to implement even a text/markdown header to request token efficient data from FQDN’s - most won’t even look to an llms.txt file and prefer Fetch
English
I do wonder if we’ll get to a point soon(ish) where agents will be much more proficient at following the AGENTS .md file.
And I keep trying to wrap my head around why they are so bad at it now.
Right now I feel like it’s a combo or context dilution and the fact that there no hard enforcement layer.
If the runtime had a more strict ruleset around AGENTS .md, i.e. if you don’t follow the rules, reject output, would that have negative consequences in other areas?
English
💯 Agents need their own discovery “UI”
Fluffless, token efficient, progressive disclosure agent ‘websites’ are a force multiplier
Agents should send and receive text/mardown headers with agentic instruction to a FQDN for further analysis and navigation
500 tokens vs 5k html and 10x faster
English
@adisingh @agentmail Your service is awesome - been using it with my agents for a couple months and it’s changed the Agent Inbox workflow
Well deserved - glad to support!
English
POV: You didn't give up.
For most of AgentMail's life, growth was flat. No one understood it but we kept pushing.. until one person did. The next 10 followed. Then, the rest.
Keep going, make that bet on your vision for the future.
Make something agents (should) want.
Garry Tan@garrytan
Sometimes startups feel like: Make people want something that they should want.
English
If just we had a standard way for integrating things between different products of the same company or even products across different companies.
Ole Lehmann@itsolelehmann
anthropic should add a simple feature to sync skills between claude chat, claude cowork and claude code and between teams i see how much people are struggling with this
English
Yeah - it’s pretty neat the way the agent reads through the available website context and you can just strip out all the fluff because an agent doesn’t want that
Structure the necessary data (price, plans, etc) and provide hierarchical linked results - agent is 10x’d in discovery efficiency (cost x time)
Can’t help but think this would be great for AEO/GEO
English
@morganlinton Ex: I built middleware to redirect agents who make a request that accepts markdown to return the markdown, progressively disclosed agent/markdown files with instructions of where to go for more details.
It’s way faster and far less expensive to navigate a website this way
English
Legit - but it’s not
Most go straight to Fetch and suck 2k tokens for a simple link/price/info - navigating human made pages and trying to make sense of the text in an unstructured way inside of rendered HTML.
Agent website UI’s can handle this - but we still need the agent to ask for it & cooperate
English
It really is - I explored this over the past few days and build out an agent ui for my app
It’s a super early space, and there’s problems, but if agents can start adopting text/markdown headers what a wonderful world it would be
Currently, llms.txt is the gateway - but we can do better
English
@masondrxy @huntlovell I’m doing something similar - hydrating MCP Endpoints/tools only when needed
Doing this with skills is a great idea - my setup uses @convex as the storage layer and api/MCP for transport. All files are embedded for easy semantic recall/search via MCP tools
English
something i've started doing for coding agents:
* keep all the skills I use in a specially formatted notion table
* have a backend that renders this table in files
* any edits I make here is immediately available to any local or cloud coding agent I spin up
English
Day 1 w/ @langChain Fleet Agents
☑️ Added MCP w/ remote FS to LangSmith via OAuth
☑️ Created a Deep Agent Specialist in Fleet from one sentence
☑️ Started work
Fleet Agent now has remote FS with Deep Search embeddings and is ready for work in less than a few mins 🚀
English
Getting started with @LangChain's new Fleet could not be easier
Here's how easy it is to setup your custom MCP servers - instantly available to any agent you create using natural language
These subtle user experiences - that's the good sauce
LangChain@LangChain
Introducing LangSmith Fleet: an enterprise workspace for creating, using, and managing your fleet of agents. Fleet agents have their own memory, access to a collection of tools and skills, and can be exposed through the communication channels your team uses every day. Fleet includes: → Agent identity and credential management with “Claws” and “Assistants” → Sharing and permissions to control who can run, clone, and edit (just like Google Docs) → Custom Slack bots so each agent has its own identity in Slack Try Fleet: smith.langchain.com/agents?skipOnb… Read the announcement: blog.langchain.com/introducing-la…
English
Double woo! v1.34 of @convex is now out that contains this! Ill make a video on the details soon!
Mikeysee@mikeysee
Woo! My biggest contribution to @convex so far has just been accepted and merged in. This one auto-injects more AI intelligence into your convex projects via guidelines and skills. Ill do a proper video on this once it lands in the CLI.
English
After I setup the MCP server in about 30 seconds -
Here's the Agent Creation Flow I used to build the @LangChain Agent using Fleet
It's just regular words, a clarifying question, and it makes the thing 🔥
The Prompt: Create an agent that is skilled at using "the MCP tools"
Git Maxd@GitMaxd
Getting started with @LangChain's new Fleet could not be easier Here's how easy it is to setup your custom MCP servers - instantly available to any agent you create using natural language These subtle user experiences - that's the good sauce
English
@Vtrivedy10 @LangChain Let me show you a demo of CR! Think Agent Cloud FS with markdown embeddings - MCP/API over any UI
English
@GitMaxd @LangChain nice!! would love to hear about the cool stuff you’re using it for and anyways we could add more :)
English