Vibe EngineeringSMIT · AI Agentic Engineering
Class 5 · 5.1

Choosing the right model for agentic coding

Class 5 is a projects day. Before we build, one decision matters more than any other: not which editor you open, but which model you put behind the wheel.

🎯 Goal: pick tool & model deliberatelyRead: 8 min🧭 Type: orientation

01Two decisions, not one

Class 4 introduced four ways to run an agent inside your editor — Cursor, GitHub Copilot, Codex and Antigravity. It is tempting to treat “which one do I use?” as the question. It isn’t. Every one of these tools is really an agentic platform: a program that calls a large language model in a loop, gives it tools, and lets it work toward a goal. Once you see that, the choice splits cleanly in two.

DECISION 1 · THE TOOLING DECISION 2 · THE MODEL The IDE Cursor · Copilot · Codex Antigravity — the workshop The LLM the brain that makes every decision +
The tools are much of a muchness. The model is where the quality — and the money — really lives.

Pick a tool you feel comfortable in and you have made a low-stakes choice; the platforms differ mostly in small ways (one reads an agents.md, another uses a .rules folder). Pick a model and you have made the decision that actually determines whether the code is any good. That is where your attention and your budget should go.

02The four tools, recapped

They feel similar because most of them are built on the same foundation — Microsoft’s open-source VS Code — and they all reach the same models.

ToolMade byWhat it actually is
CursorAnySphereA fork of VS Code — its own application that looks and feels like the editor you know.
GitHub CopilotGitHub / MicrosoftAn extension loaded into the real VS Code, adding an agent chat panel.
CodexOpenAIUsable as a VS Code extension, but most commonly driven from the command line (CLI) — which is where next week goes.
AntigravityGoogleAnother standalone fork of VS Code, in the same spirit as Cursor.
Tooling ≠ model

The IDE is the workshop; the LLM is the worker you hire into it. Most tools can hook up to most models, so switching tools rarely changes your results much — switching models changes everything.

03The models behind the agents

Whether you noticed or not, Class 4 quietly used several different models. On “auto” mode a tool picks for you and often won’t say which it chose. Here are the ones in play, grouped by tier.

ModelMakerTierContext window
ComposerAnySphere (Cursor only)Fast frontier~200K
Claude Haiku 4.5AnthropicFast frontier~200K
Claude Sonnet 4.5AnthropicMid / strong~200K
GPT-5.2 CodexOpenAITop frontier~272K
Gemini 3 Pro / FlashGoogleTop frontier / fast~1M
Claude Opus 4.5AnthropicTop frontier~200K
Fast frontier

Leading models from the small-and-quick camp — Composer, Haiku 4.5. They generate code fast and cost little, but need a closer eye.

Top frontier

The strongest models available — GPT-5.2 Codex, Gemini 3 Pro, Claude Opus 4.5. Slower and pricier, far more reliable on hard, multi-step work.

04Three rules of thumb for picking a model

Favour smart over fast

A strong model that takes longer but gets it right beats a quick one that produces something broken you then iterate on. Intelligence over speed usually reaches a faster finish.

Set a budget, then buy the most intelligence it allows

Trying to save money with a cheaper model tends to cost more — extra iterations, wasted time. Decide the budget for a task, then pick the strongest model that fits it.

Smaller model → bigger spec, more oversight

Free and open-source models can work, but only if you write far more precise prompts — think of it as authoring a detailed spec — and supervise closely.

05So — when is it safe to YOLO?

“YOLO” means kicking the agent off and letting it run with little supervision. Whether that’s wise depends almost entirely on the tier of model you’re trusting.

Which tier? Fast / small model Don’t YOLO Fast frontier Step-by-step, watch the diffs Top frontier YOLO can work well The risk with small models isn’t mainly safety — it’s coming back to a pile of nonsense and losing the time.
Match your supervision to the model, not to your mood.
Watch the edits

With a fast frontier model, work in baby steps and keep your eyes on the diffs as it changes code. Reserve true hands-off YOLO for the top models — GPT-5.2 Codex, Gemini 3 Pro and, the instructor’s favourite, Claude Opus 4.5.

Key takeaways

  • Two decisions: the tool (low stakes) and the model (where quality and cost live).
  • The four IDEs are agentic platforms — a loop that calls an LLM with tools — and most reach the same models.
  • Favour smart over fast, set a budget then buy the most intelligence it allows, and give smaller models bigger specs.
  • Only YOLO with top-frontier models; supervise fast models step-by-step with your eyes on the diffs.