Inside the Claude Code Obsession

A healthy signal that a product has survived the hype cycle is when conversations shift from awe to utility. Users stop discussing what a tool could do and start trading notes based on demos—swapping guidelines, hacks, and niche use cases the creators likely never anticipated.

And even though Anthropic’s agentic coding tool Claude Code reached $2.5 billion in annual revenue run rate, engineers in Silicon Valley and elsewhere have been quietly experimenting with it and integrating it into their workflows. Perhaps this abstraction is the clearer signal of its maturity.

A quick search on X shows developers getting straight to business and including it in their tech stack. Many feel it marks a landmark moment in software development.

Parallel Subagents and Context Decomposition

Most coding assistants still behave like a single reasoning engine attempting to ingest an entire repository at once. Claude Code instead decomposes work into parallel threads, each responsible for gathering its own context and reporting back.

Calvin French-Owen, an entrepreneur-in-residence at Thrive Capital and a former software engineer at OpenAI, highlighted that subagents in Claude Code help manage complexity.

“I think one of the things that Claude Code does in particular that's really amazing is split up context well,” he said in a recent Y Combinator roundtable.

Subagents partition context into focused roles, allowing each agent to operate within a scoped slice of the problem rather than sharing overloaded prompts, with an orchestrator aggregating their outputs into a coherent result.

Under the hood, many of these subagents run on lighter-weight models such as Claude Haiku, which makes it cheap and fast to spin up multiple parallel workers. Each subagent can independently analyse files, propose changes, or run small reasoning loops, while the main Claude model focuses on higher-level synthesis and final decision-making. The result is better context, faster iteration, and higher-quality solutions on large codebases without flooding a single model with everything at once.

Rather than competing purely on single-model reasoning strength, leading systems are beginning to rely on an orchestration agent that sends out copies of itself to work on pieces of a problem, then synthesises the results. This design changes the developer’s role—humans specify intent and evaluates outcomes, the system handles navigation, exploration, and execution.

Why Claude Code is Not an IDE

French-Owen argued that Claude Code’s effectiveness lies in being a terminal-first tool that operates outside the traditional IDE (integratedf development environment). “I think it's important that it's not in an IDE because it distances you from the code that's being written,” he said.

Traditional IDEs involve manually navigating files and mentally tracking system state. A command-line interface (CLI) frees the tool from that file-by-file paradigm and enables a more orchestrated experience.

“When I'm using Claude Code, I feel like I'm flying through the code,” French-Owen said. “There are all sorts of things going on—little progress indicators, status updates, but the code that's being written is not the front and centre thing.”

Claude Code presents a plan before editing files and exposing diffs (comparing files line-by-line), allow for intervention mid-task. That workflow mirrors how experienced engineers already operate in production environments.

French-Owen also noted that Claude Code’s bottom-up, CLI-first distribution is an important advantage. It can be downloaded and deployed locally without organisational approval. This is condusive for experimentation and allows usage patterns to emerge organically, rather than being dictated by top-down enterprise rollout.

What Developers are Actually Using it for

AIM reached out to multiple developers who use Claude Code extensively.

Pranshu Agrawal, a lead AI engineer at a global enterprise AI and analytics company, says control is the defining factor. “When I'm using Claude Code, it feels like I'm actually driving the process because I can pause it, I can redirect it or change the scope without losing any context.” 

This contrasts with vibe coding tools that often prioritise rapid generation over inspectability, leaving developers with limited ability to intervene, steer execution, or understand how decisions are made.

Claude Code’s plan-first workflow also helps him prevent mistakes. “It also has a workflow-like setup on its own. It proposes a plan, has checkpoints, and then executes,” he explains.

In large enterprise repositories, Agrawal says Claude Code selects files conservatively and respects architectural boundaries. “It respects the architecture that I've been working with, which is crucial in a production setup.”

Beyond control and architectural boundaries, some engineers focus more on inspectability and early failure detection.

Kirk Kaiser, a software engineer and author of Make Art With Python, considers inspectability essential for production use. “The ability to use plan mode and give feedback via tab for both ‘yes’ and ‘no’ lets me iterate on creating a plan, read through the strategy and approach, and understand how much total work Claude would take on, which helps me guesstimate the likelihood of success,” he tells AIM. 

He relies on diff-level feedback to catch failures early. “Having planning mode, diffs, and the ability to provide feedback on each change lets me see when Claude is going off the rails or doing something in a way that isn’t going to work.” 

Still, Kaiser stresses humans cannot shrug away responsibility. “I think we've got to understand our code if we're going to deploy it in production, and that means reading through and understanding both the LLM's approach and its implementation.”

Meanwhile, Claude Code’s capabilities are expanding beyond pure coding through the model context protocol (MCP) and Skills functionality.

Aditya S Kolavi, an AI researcher at CognitiveLab who has been using Claude Code for over six months, says the tool has improved significantly over competitors like Cursor, Copilot, and Windsurf.

“With the growth of MCP and Skills, it has evolved into more than just a coding assistant. I now use it for several non-coding tasks as well, such as managing my calendar and handling various automation workflows,” he adds.

On the production side, he highlights performance inside structured codebases. “When there is a proper codebase with good structure and verifiable test cases, it adapts quickly and can add features efficiently. It is also capable of handling large codebases effectively and making changes across the codebase without much friction.”

While Claude Code, with the latest frontier model Opus 4.6, still shines in undo/rewind commands, iterative vibe coding, frontend work, and shader effects, its main rival isn’t standing still. 

OpenAI’s GPT-5.3-Codex has upgraded from a focused agentic coder into a broader general-purpose assistant, adding real-time interactive workflows, native multi-agent orchestration, and the new Codex macOS app for parallel agents, long-running tasks, and direct Git integration.

Developers now see Codex rapidly closing Claude’s earlier leads in steerability and undo-style editing, while pulling ahead on large codebases, boilerplate generation, and efficiency.