Ben

My passion is solving interesting problems with interesting people using the latest AI tools and approaches.

[Ben @BenHarrell_]

@CharlesTingler pretty straight forward, personally I wouldn’t include temple system. I interpret it as the old covenant as the thing that passed away. maybe that’s what you meant ?

[Ben @BenHarrell_]

Coding agents expose a hard truth about evolution: optimization does not converge on what’s good, only on what survives the pressures we define.

@LangChain

[Ben @BenHarrell_]

@deusimpera must be Macca-bees. 😛 @Preteristesch

[Ben @BenHarrell_]

@WassonWatch Oh no the Assyrians are back! @AmericanVision

[Ben @BenHarrell_]

@mattpocockuk imagine if @jason turns out to be correct on small / OS blockade #Iranthropic ?

[Theo - t3.gg @theo]

My Nest fire alarm expired. Apparently they don’t make them anymore. It’s beeping very loudly once a minute. It’s 20ft up on my ceiling.

I am very very annoyed right now.

[Ben @BenHarrell_]

@davis7 wait until you see 5.5! 🫠

[Ben @BenHarrell_]

Naming your variables MyClass, MyObject, and doThing() probably boosts AI coding performance by like 37%… just matching its training data 😅

Can someone confirm? @karpathy

[Ben @BenHarrell_]

@apivault_labs @mattpocockuk have another call of course! 🤣

[Ben @BenHarrell_]

@hwchase17 wouldn’t the langchain pattern be to support all via integrations?

[Ben @BenHarrell_]

I’m having 80s flashbacks now when coding….anyone remember King’s Quest 🤣

[Ben @BenHarrell_]

@jamonholmgren SETI - search for extra token inference 🤣 bring back the screensavers!

[Ben @BenHarrell_]

Using a sub-agent in #langchain feels a lot like spawning a subprocess in Python.

synchronizing state or memory between them becomes an interprocess communication problem.

That’s why the #deepagents instructions matter so much—they are basically the IPC layer for agents.

[Ben @BenHarrell_]

@mick__net @masondrxy @hwchase17 mcp_client = MultiServerMCPClient({
"my_server": {
"transport": "stdio",
"command": "python",
"args": ["my_mcp_server.py"]
}
})

agent = DeepAgent(
name="",
llm=llm,
tools= mcp_client.get_tools()

[Ben @BenHarrell_]

@mick__net @masondrxy @hwchase17 just load them as tools using the client and pass to deepagent constructor?

[Ben @BenHarrell_]

@SkullyTougs @palmettoarmory we need a chant like:

Ex five seven
Ex five seven

🤣

[Ben @BenHarrell_]

@mattpocockuk and most of the common frameworks give you sub agents for free when you do

[Ben @BenHarrell_]

@davis7 is rxjs a good fit?

use Observables to orchestrate the HTTP stream (fetch → compose → transform), then bridge to an async iterator when you need pull semantics. Rx for flow control, for await for chunk-by-chunk iteration. Push core, pull at the edge.

[Ben @BenHarrell_]

lately i’ve been trying to map knowledge ai to physical ai and i keep coming up with weird conclusion like:

teachers might be the new software engineers

imagine the robotic equivalent of a phys therapist teaching them to interact with human world 🤨

[Ben @BenHarrell_]

@mattpocockuk maybe the future in this is live docs lazy loaded? more like JiT than cache?