Agentic Coding Explained

Agentic coding is the style of programming where you describe an outcome and an LLM-driven agent executes the full loop: reading files, editing them, running tests, and looping until the work is done. It sits between autocomplete and full autonomy.

How it differs from autocomplete

Classic AI autocomplete — GitHub Copilot's ghost text, for example — produces the next few tokens based on cursor context. The human stays in the loop for every change.

Agentic tools invert that. You write a goal ("add pagination to the users API and update tests"), and the agent:

1. Searches the repo for relevant files.
2. Drafts changes across multiple files.
3. Runs pnpm test or cargo check to verify.
4. Iterates on failures before showing you a diff.

The current tools

• Cursor — VS Code fork with an agent mode that edits across the workspace.
• Claude Code — Anthropic's terminal-first agent; strong on long-running tasks.
• Windsurf — Codeium's agent-native IDE, formerly Codeium Cascade.
• Cline — open-source agent that runs inside VS Code.
• Aider — git-aware CLI agent, predates the current wave.

Trade-offs and failure modes

• Review overhead. A 400-line diff from an agent still needs line-by-line review. The time saved in typing is partly spent in reading.
• Context drift. Long sessions accumulate stale context and start regressing previous fixes. Start fresh chats for independent tasks.
• Tool blindness. Agents can't see things outside the files they read — runtime config, CI state, production logs. They hallucinate around those gaps.
• Cost. A single feature can burn several dollars of API calls. For hobby work that adds up fast.

When NOT to use an agent

One-line bugs, highly sensitive code (crypto, auth), and changes that require product judgment are faster and safer to do by hand. Save the agent for tasks with clear test signals and moderate scope.