I thought claude already worked that way. lmao     See Less

Learn english first pls     See Less

The example you gave of getting the emails, shows the function returning IDs and then for each ID returning the emails. This is unnecessary. You could have written a function that doesn'     See More Ds but just the emails. The looping on IDs would take place in the function. You failed to explain the real benefit of using the code execution tool. It lets you run terminal commands in a sandboxed environment and create and edit files. Your video title is false. Anthropic hasn't "killed tool calling". Read the official docs on Claude's site and you won't find "deprecated" or suggestions to switch to the code execution tool. Your video is like so many. Too much fluff to get views but not grounded in reality.    See Less

@grok summarize this     See Less

Isn't this from November 2025?     See Less

I mean, look at pi, it’s the agent inside clawdbot and it doesn’t support mcp at all. Bash is all you need. Llm is great at writing bash commands and then linking them together with pipe     See More cript.

This isn’t new and we don’t need a new tool thing for it, just bash, we already have tons of ash tools that can filter and compose together to make complex data pipelines without using up context    See Less

I'm working on a Two-Model prototype, one big brain (like GPT 5.2) handles the thinking, then gives instructions to a small brain(like Qwen3-8b with mcp skills) which runs localy on my     See More h only the essential tonkens, the small brain handles all the tool-calling accordingly.
So the big brain be like: "Hmm... user mentioned he will be attending his son's school event, but is that going to be a conflict to his work schedule? Small brain give me a rundown of his work schedule from his Google Calender on XXX.""
The small brain be like:"Capiche! 1 minute! <running gog shit for 4 thousand tokens....>, Hey big brain ,you are right, his schedule is full on XXX but is open on YYY"
Big Brain:"Thanks Small Brain you are the best. Hey User, guess what, bad news...."
I'm even considering introducing a third barin which is the friendly brain with strong role-play ability and low price. So "thinking", "Talking", and "Action" will be handled by 3 different model and system-prompts, Hope this design can address the attention drift caused by excessively long prompts and the Turing-damned high token costs.    See Less

Anthropic's Programmatic Tool Calling fundamentally shifts agentic behavior by allowing the LLM to generate and execute code (like Python) internally within its thought process, rather t     See More creating external CLI commands. This enables the agent to orchestrate multi-step tool calls, use loops and conditionals, and crucially, apply dynamic filtering to tool outputs (e.g., precise log grepping) before consuming valuable context window tokens. This innovation significantly reduces redundant round trips, enhances deterministic execution, and prevents "context window bloat" from noisy tool results, making agentic workflows much more efficient and accurate, especially in complex, recursive delegation chains.    See Less

Is this basically a form of RLM ?     See Less

Scamthropic     See Less