Agentic coding is a software development paradigm where an AI agent autonomously reads, plans, writes, tests, and iterates on code — executing multi-step workflows rather than suggesting individual lines. The term was coined by @Anthropic to describe the capabilities of @Claude Code, distinguishing it from the autocomplete-style AI coding that preceded it.
How It Differs from AI-Assisted Coding
Traditional AI coding tools (GitHub Copilot, Cursor's autocomplete, early ChatGPT) are reactive: the developer writes, the AI suggests. The human stays in the driver's seat, accepting or rejecting line-by-line recommendations. The context is usually a single file. Agentic coding inverts this. The developer describes intent. The agent drives: reading the full codebase, understanding the architecture, planning the implementation, writing code across multiple files, running tests, interpreting failures, and iterating until the task is done. The human provides direction and reviews output. The agent handles execution. This mirrors the broader shift in @human-AI augmentation — from the AI as a tool you operate to the AI as a colleague you direct. The skills that matter change: from writing syntax to describing systems, from debugging line-by-line to reviewing architecture, from typing fast to thinking clearly.
What Makes It Work
Codebase Awareness
Agentic coders read entire repositories — file structures, dependencies, test suites, configuration files, and @CLAUDE.md instructions. This full-codebase context is what enables multi-file edits that maintain consistency. A tool that only sees one file can't refactor safely.
Planning Before Acting
Claude Code's plan mode is the clearest example: before writing code, the agent reads the codebase, proposes an approach, and waits for approval. This plan → execute → verify loop produces better results than jumping straight to code generation, especially for non-trivial changes.
Tool Use and Shell Access
Agentic coders don't just write code — they run it. Shell commands, test suites, build tools, git operations, and @MCP-connected external services are all available. The agent can verify its own work, which creates a closed feedback loop: write → test → fix → verify.
Sub-Agent Orchestration
At scale, a single agent context window isn't enough. @Multi-agent orchestration — spawning sub-agents for parallel tasks, delegating research to background agents, coordinating across files with locks — is what makes agentic coding viable for production systems. @BrianBot's 57-agent ecosystem runs primarily on agentic coding patterns.
Why Anthropic Coined It
Anthropic needed a term that separated Claude Code from the autocomplete tools it competed against. "AI coding assistant" implied the old paradigm — the AI assists, the human codes. "Agentic coding" implies the new one — the AI agents, the human architects. The term stuck because it described something genuinely different that practitioners recognized from their own experience. The 300% growth in Claude Code usage since the Claude 4 launch suggests the paradigm is resonating. Developers and non-developers alike are discovering that describing what you want and reviewing what the agent builds is a fundamentally different — and for many, more natural — way to create software. I didn't know there was a name for how I work until Anthropic coined this one. But "agentic coding" is exactly right. I don't write code. I describe systems and review what Claude Code builds. The agent reads my codebase, understands the architecture from my @CLAUDE.md, plans the implementation, and executes. I steer, verify, and decide. The shift from "AI helps me code" to "AI codes and I architect" was gradual. It started with trusting Claude Code for small tasks — a function here, a refactor there. Then whole features. Then entire applications. The trust compounds the same way the @Augmentation Stack compounds — each successful execution earns the next delegation. Now I think of coding the way I think of managing: the goal isn't to do everything yourself. It's to create systems where the right work happens without you doing it manually. That's what agentic coding is — management applied to software creation.
Contexts
- #agentic-augmentation
- #anthropic
- #claude-code