Skip to content
#Agents

How Claude Code Works - How Claude Code Works Inside: An Architecture Discussion by PromptLayer Founder Jared Zoneraich (Agents column)

#Agents #AI #ML #Software #Engineering #Architecture

I saw it. note from the AI Engineer Conference in New York, which was made by Jared Zoneraich, the founder PromptLayer And a man who sees thousands of prompts a day. Its platform handles millions of LLM requests, and the team has rewrote its engineering organization around Claude Code. Jared is not from Anthropic, but his insights are the result of dogfooding and analyzing the work of thousands of developers. Jared’s main point is that coding agents didn’t work because of complex architecture, but because of its simplification. Instead of RAGs, vector bases, and complex orchestrators, Anthropic went down the path of “give models tools and don’t get in the way.” In more detail, the author’s theses look like this 3.

✈️ Architecture = one while-cycle + tools

# n0 master loop
while (tool_call):
    execute_tool()
    feed_results_to_model()

Okay. No more branches, subgraphs and state machines. The model decides what to do next. This is the N0 loop inside the Claude Code.

🤖 Tools copy developer behavior in terminal

  • Bash - the king of all instruments. The model can create a Python script, run it, see the output, delete it. This gives the flexibility of thousands of utilities without custom development.
  • Read/Grep/Glob - the search you'd be looking for. No vector bases, just grep and glob patterns.
  • Edit Diffs instead of overwriting files. Faster, cheaper on tokens, fewer errors.
  • Todos Structured planning through prompt rather than deterministic code.

🗒 Todo-lists work on an honest word The system does not force the execution of tasks deterministically. Instead, the system prompt inserts the instruction "one task at a time, mark completed". The model just follows the instructions – and it works because modern LLMs understand the context well.

🍬 Context Management with H2A and Compressor

  • H2A (Half-to-Half Async) - double buffer for pause/reopening. You can intervene mid-task without restarting.
  • Compressor wU2 - It works on ~92Percent of context filling, sums up the middle, leaving the beginning and the end. This gives the model a "place to think" before the crisis.

💯 Simplicity> complexity According to Zen of Python, “Simple is better than complex.” Complex is better than complicated”. All attempts to protect the model from hallucinations through complex scaffolding is a technical duty. It is better to wait for the model improvement and remove the extra code.

In the end, all this can be reduced to simple advice:

  1. Stop over-optimizing.. If you build agents and write crutches to work with current models, you’re wasting time. It is better to invest in clean and simple architecture.
  2. Bash as a universal adapter. Instead of writing custom tools for each task, give the agent access to the shell. All utilities (ffmpeg, git, grep) It's already in the system.
  3. Prompt engineering > complexion. CLAUDE.md file with instructions is more efficient than local vector databases. The model explores the repository itself if it knows what to look for.
  4. Prepare for the next wave. If your team hasn’t rewrote workflow around coding agents yet, you’re falling behind. PromptLayer made the rule: "if the problem < 1 Do it with Claude Code, don’t plan.

As Jared says, “Less scaffolding, more model.”

P.S. This is exactly what Nik Pash, Head of AI, said. Cline reportHard Won Lessons from Building Effective AI Coding Agents told

#AI #ML #Agents #Software #Engineering #Architecture