The SPACE Framework for AI Code Generation

The SPACE framework breaks down AI-assisted development into five dimensions: Search, Plan, Act, Critique, and Evaluate. It gives teams a shared vocabulary to discuss where AI helps and where it hurts. Here is how to apply it with today's agentic tools, and where AZMX AI fits into the stack.

Why the SPACE framework matters in 2026

In 2024, most developers treated AI like a fancy autocomplete. By 2026, agentic tools can write entire modules, refactor codebases, and even run terminal commands. The problem is that trust hasn't caught up to capability. Developers needed a mental model to reason about where AI adds value versus where it introduces risk. The SPACE framework—popularized by researchers studying developer productivity—fills that gap.

SPACE stands for Search, Plan, Act, Critique, and Evaluate. Each dimension corresponds to a phase in the AI-assisted development workflow. Understanding these phases helps you pick the right tool for the right job, set appropriate approval gates, and avoid common pitfalls like hallucinated dependencies or leaking secrets.

Search: Finding context in a massive codebase

The first dimension, Search, is about locating relevant code, documentation, and patterns. Before an AI can generate useful code, it needs to understand the existing codebase. This goes beyond a simple grep. The AI needs to trace function signatures, understand project conventions, and identify the correct files to modify.

Most AI coding tools today embed the whole workspace into a vector index. Some, like Cursor and Windsurf, maintain a persistent index that updates as you edit. Others, like Claude Code and Aider, rely on the developer to explicitly add files to context. The difference is important: persistent indexing saves time but can miss recent changes, while explicit context is more accurate but requires manual effort.

For teams working on large monorepos, the Search phase often becomes the bottleneck. A good practice is to keep a project-level AZMX.md file (or similar) that documents architecture decisions, known patterns, and files the AI should reference. This acts as a static context map that any tool can read, reducing the chance of search failures.

Plan: Structuring the work before writing code

The Plan dimension asks: what exactly should the AI do, in what order, and what if something goes wrong? This is where agentic frameworks shine—or fail. A naive agent might go straight to editing files without verifying assumptions, leading to cascading errors.

Tools like Continue and Cline implement a plan-then-execute loop: the agent writes a plan, shows it to the developer for approval, then acts. Other tools like Codeium's agent skip this step and produce faster but riskier outputs. The SPACE framework recommends always including a plan step for any change that affects more than one file or touches production code.

In AZMX AI, the agent writes a detailed plan before making any edits. The developer can reject, modify, or step through each change using per-hunk diffs. This gives the human full control over the Act phase while still benefiting from AI speed. The plan is also stored in the session log, making it auditable later.

Act: Executing changes with precision

Act is where the AI actually modifies files, runs commands, or interacts with the terminal. This is the most visible phase and the one most developers associate with