Magic

Magic is developing its own frontier models trained specifically for software engineering, with a focus on extremely long context windows so an agent can reason over an entire large codebase at once. It’s earlier-stage and less publicly available than most tools in this list, but notable for the scale of investment behind an autonomous-engineer approach.

Why it stands out

• Long-context-first architecture, betting that reasoning over a whole codebase in context beats retrieval-based approaches for complex changes.
• Builds its own models rather than wrapping an existing LLM API, giving it more control over code-specific training.
• Heavily backed research effort, with major infrastructure investment aimed at training models purpose-built for autonomous engineering.

Good for

Teams and researchers tracking the frontier of long-context, autonomous coding models rather than looking for a tool to use today.