Definition

Agentic coding

Agentic coding is software development where an LLM-powered agent plans, edits, runs, and verifies code on its own using tools, not just autocomplete.

Agentic coding is the practice of letting an LLM-powered agent take multi-step actions on a codebase — reading files, editing them, running tests, checking output, and iterating — instead of returning a single completion. The agent decides what to do next based on what it observes, closing the loop between "suggest" and "do." Tools like Claude Code, Codex CLI, Qwen Code, and Aider are all agentic.

Why it matters

Traditional autocomplete tools (early Copilot, plain ChatGPT pasted into an editor) leave all the orchestration to you: you run the tests, you apply the diff, you tell it what failed. Agentic coding moves that loop inside the tool. You describe the goal — "fix the flaky auth test" — and the agent investigates, patches, runs the test, and re-patches until it passes.

This changes how developers structure work. Instead of one session per prompt, you often want several agents running in parallel on different branches or tasks. SpaceSpider exists specifically for this: it hosts a grid layout of agentic CLIs in one window so you can watch several autonomous agents work at once. See getting-started.

How it works

An agentic coding session is a loop: the agent receives a goal, proposes a tool use call (read file, run shell command, apply diff), the CLI wrapper executes it, and the result is appended to the context window as observation. The LLM reads the observation and emits the next tool call. Loop until the goal is met or the agent hits a stopping condition.

Most modern agentic CLIs add plan mode for upfront strategy, checkpoints to roll back changes, hooks to customize behavior, and sandboxing to limit blast radius. The Model Context Protocol is emerging as a shared way to plug external tools (databases, APIs, browsers) into any agent.

How it's used

Typical agentic coding tasks:

  • "Migrate this file from Jest to Vitest"
  • "Find and fix all N+1 queries in the orders service"
  • "Write a failing test that reproduces issue #482, then fix it"
  • "Refactor this component to use the new design system"

Running several of these in parallel panes — one per task, each in its own git worktree — is a common SpaceSpider pattern. See /compare/claude-code-vs-cursor for how agentic CLIs differ from editor-embedded AI.

Related terms