# Anatomy of Autonomous Coding Agents - Public Source Note

Date: 2026-02-22  
Status: public scrubbed note  
Prepared by: Greyforge Labs

Related chronicle: https://greyforge.tech/chronicles/anatomy-of-ai-coding-agents  
Primary public corpus: https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools

## Public Scope

This note backs the chronicle that compares the visible harness shape of major coding agents.

Greyforge reviewed public prompt and tool snapshots for Cursor, Windsurf, Devin, v0, Lovable, Replit, and related systems in the linked corpus.

The claim is not that prompt snapshots reveal every behavior of every product.
The claim is narrower: they reveal enough about authority boundaries, memory posture, tool routing, and review structure to support an architectural comparison.

## Public Findings

1. Most systems still concentrate planning, execution, and review inside one primary loop.
2. Memory and tool depth vary, but role separation remains thin.
3. When subagents appear, they are usually utility workers or explorers rather than independent quality gates.
4. The strongest improvements are harness refinements inside a single-agent frame, not durable multi-role collaboration.

## Why That Matters

Greyforge's comparison argument is about architecture, not branding.

A stronger tool catalog or a longer prompt does not automatically create separation of concerns.
If one loop still has to plan, execute, self-review, and manage the conversation at once, the system keeps the same structural blind spots.

## Limits

- prompt snapshots can lag product changes
- public corpora can miss hidden runtime behavior
- vendor products may have internal systems not visible in the prompt layer

Those limits do not erase the architectural signal. They only narrow the strength of the claim.

## Greyforge Public Position

The prompt corpus shows a market that has improved its harnesses without fully solving authority separation.

Greyforge treats that as evidence for a broader thesis:
better autonomous engineering systems need bounded roles, independent review surfaces, and durable memory discipline, not just a smarter single loop.