Coding Agents

1. Install Ollama → ollama pull qwen2.5-coder:14b (~9 GB — minimum viable for coding agents).
2. pip install aider-chat (Aider — the leading open-source local coding agent).
3. cd /path/to/your/repo && git init (Aider needs a git repo to track changes).
4. aider --model ollama_chat/qwen2.5-coder:14b — opens the aider TUI.
5. Ask: "Create a REST API endpoint for user registration with email validation, password hashing (bcrypt), and JWT token return. Use Express.js." Aider: reads the repo → creates routes/auth.js → creates models/User.js → installs bcrypt + jsonwebtoken → writes tests → runs tests.
6. First agentic task completion in 2-10 minutes depending on complexity.
7. For VS Code: install Cline extension → configure Ollama → use DeepSeek Coder V3. Cline reads files, writes code, runs terminal commands, and iterates on errors.
8. Key: coding agents work best when the task is well-specified. "Build a todo app" fails. "Create Express.js CRUD API for todos with MongoDB, input validation, and error handling" works.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs aider with Qwen 2.5 Coder 14B at 25-35 tok/s — handles multi-file tasks (CRUD endpoints, refactors, test writing) on repos up to 50K lines. Each agent step (read file, think, edit, run test) takes 5-15 seconds. A 5-step task completes in 1-2 minutes. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe. Total: ~$400-480. For coding agents specifically, 14B is the minimum for multi-file edits. 7B models get lost across files. $400 gets you a capable coding agent for small-to-medium projects.

Used RTX 3090 24 GB ($700-900, see /hardware/rtx-3090). Runs aider/Cline with DeepSeek Coder V3 at 15-20 tok/s or Qwen 2.5 Coder 32B at 35-50 tok/s — these models handle complex multi-file architectures, database migrations, and integration tests across repos up to 500K lines. For professional developers using AI pair programming daily: the 32B class of models reduces the "fix my own fix" cycle by 50% compared to 14B. Total: ~$1,800-2,200. For the fastest iteration: RTX 4090 ($2,000) + Qwen Coder 32B at 60-80 tok/s — near-instant code generation.

The mistake: Letting the coding agent run autonomously for 20 minutes on a complex refactor, then discovering it deleted features, duplicated code, and introduced circular imports — and git reset won't help because it committed 15 times. Why it fails: Coding agents compound errors. Step 1: makes a small mistake. Step 2: "fixes" the mistake with a workaround that breaks something else. Step 3-15: builds on broken foundations. After 20 minutes, the codebase is a house of cards. The agent is a junior dev without supervision — it needs review at each step. The fix: Use coding agents incrementally. Ask for one small change at a time. Review the diff. Run tests. Commit. Then ask for the next change. This is slower per-step but 10× faster overall because you avoid the compounding-error cleanup. Aider's /undo command helps but only within a session. For Cline, review every file edit before approving. Coding agents are pair programmers — not autonomous contractors.