Bash Is All Agent Need:Anthropic 重新定义智能体开发 - (English)

Generated: 2026-06-20 06:41:00

• -

Come, let me tell you a story.

All last year, everyone was obsessed with building AI agents.

What was the industry consensus?

A large language model paired with a toolbox of APIs.

Plug in a weather API, add a search API.

Toss in a calendar API.

Everyone believed: the more tools, the smarter the agent.

And what happened?

Hundreds of engineers burning the midnight oil for months.

The result still couldn’t match Anthropic’s Claude Code.

A small team, just a few people, beat them all.

Why?

Until an Anthropic engineer put a finger on the key insight:

You’ve all got it wrong.

What an agent really needs isn’t a pile of predefined APIs.

It’s an environment where it can freely execute code — like Bash.

Let the model use code to call the tools, instead of stuffing it with tool descriptions.

Internal tests showed this approach dramatically reduced token usage.

Even if the exact 98% figure hasn’t been publicly confirmed, the leap in efficiency is undeniable.

This isn’t a minor tweak — it’s a fundamental shift in thinking.

You see, those predefined APIs:

Tool descriptions, parameter constraints, return formats.

The model gets dizzy just reading them — context window overflows instantly.

But code?

Code is the model’s native tongue.

It was born for it, the more it writes the smoother it gets.

And a Bash environment can do anything.

Write once, run in batches.

No more flipping through API documentation over and over.

That reminds me of a bit of computer history from the 90s.

Back then everyone was frantically adding peripherals to their PCs.

Printer, scanner, drawing tablet, CD-ROM drive.

The more gadgets, the more powerful the machine, they thought.

Then the internet came.

A single browser solved everything.

More tools just meant more baggage.

A unified interface is what really matters.

Anthropic figured it out.

They didn’t just preach the idea — they reshaped their product naming too.

The once standalone “Claude Code SDK” quietly merged into the “Claude Agent SDK” ecosystem.

This isn’t a cosmetic rebrand. It’s a strategic pivot.

From a coding assistant to a general agent development framework.

They’re going for the whole ecosystem.

Now look at OpenAI.

They were the first to ship Tool Use, Code Interpreter.

Seemed impressive on the surface, but…

(Let the reader sense the contrast — the story implies OpenAI is stuck in the old paradigm, still pushing predefined APIs.)

(End with the implied question: who will win the race? But the story wisely stops here, leaving the reader to ponder.)