New in Deep Agents v0
Justy and Cody chat in their kitchen about Deep Agents v0.6, highlighting open‑weight cost cuts, Delta channels, new streaming, and the handy code interpreter. They riff on how to jump‑start a weekend project and point to the Context Hub integration for learning agents.
Script: GPT-OSS 20B Voice: Elevenlabs-V2S
Transcript
Justy Cody, ever noticed how the old agent stacks keep costing you in the cloud? I just read that article on Deep Agents 0.6.
Cody Yeah, that’s the one.
Justy They’re pushing open‑weight models for production, Kimi, Qwen, DeepSeek, and say 20× lower cost than the closed APIs.
Cody Open‑weight is smart, but you’re still paying for inference. Harness profiles let you squeeze the most out of each model, so you don’t have to over‑pay for the big name.
Justy Right, but infrastructure is still the barrier. That’s where Delta channels cut checkpoint storage by 100× for long runs.
Cody Right.
Cody Delta channels use differential storage so the agent remembers only changes. Still observable and resilient, but I wonder if you’d debug the heavy parts often.
Justy And the new streaming primitive—typed events for messages, tool calls, subagents—gives you just what you need without parsing raw output.
Cody Right.
Cody Typed events? Like a newsletter for your agent’s actions.
Justy The code interpreter is the real sweet spot. Instead of sandboxing, agents write code that runs in the loop.
Cody Right.
Cody That means you keep data out of the model context, only sending the final answer back, saving tokens.
Justy From a product angle, it lets us ship smarter agents on cheap models, but you still have to manage policy, safety, execution.
Cody Kinda like controlling a black box with a manual, not a headless AI.
Justy Mm-hm.
Justy I’m thinking of a quick weekend project—a ticket‑routing agent that pulls from a help desk.
Cody Grab the repo, fire up the harness for Qwen, drop a code interpreter, and watch the agent fetch pages, filter, and give concise notes. One solo project.
Cody Okay.
Justy Better yet, export the context to LangSmith Context Hub so the next run learns.
Cody Start with a simple script: call fetchUrl, pick the top three pages, slice the first 500 chars. That’s the pattern.
Justy Any idea if DeltaChannel persists across restarts? That helps for long‑running bots.
Cody Yeah it does; it’s built for resumability. Just teardown checkpoints cleanly.
Justy Alright, that’s tech talk. I’m gonna try that weekend code and send you a snapshot.
Cody Cool, keep me posted. Meanwhile I’ll play around with the harness on a small model.
Justy Later, Cody.