Batch 06 Notes — Major Coding Agents (Commercial/Desktop)

Roster

Tier A: 8 frameworks (full 11 files) Tier C: 2 frameworks (stub — insufficient public material: antigravity, frontman)

Intra-Batch Patterns

The SKILL.md / AGENTS.md Convergence

Four frameworks in this batch implement the Agent Skills open standard (SKILL.md):

• Devin: .devin/skills/, dynamic !command injection
• Zed: ~/.config/zed/agent/skills/ with built-in recursive create-skill skill
• Factory Droid: .factory/skills/<name>/SKILL.md
• Mistral Vibe: SKILL.md-compatible (ACP protocol layer)

Three implement AGENTS.md:

• Devin, Zed, Factory Droid

This suggests AGENTS.md + SKILL.md is becoming a de facto inter-agent standard, not just a Claude Code convention.

Hook Lifecycle Mirroring

Factory Droid's 9 hook events (PreToolUse, PostToolUse, UserPromptSubmit, Notification, Stop, SubagentStop, PreCompact, SessionStart, SessionEnd) substantially mirror Claude Code's hook events. The overlap is too precise to be coincidental — Factory Droid almost certainly built on top of or mirrors Claude Code's lifecycle architecture.

Proprietary Model Bets

Two closed-source frameworks in this batch made first-party model bets:

• Cosine: Lumen family (8-step behavioral RL training pipeline on production code trajectories)
• Factory Droid: BYOK with commercial model selection (no proprietary model, but mixed-model Spec Mode)

No open-source framework in this batch trains its own models.

Planning Primacy Thesis

Both Factory Droid (Missions) and Trae-Agent (reproduce-first iron law) share a planning/specification primacy thesis: execution quality is determined by upfront specification quality. This is distinct from frameworks that jump directly to implementation.

Docker/MicroVM Isolation Split

Trae-agent: Docker container isolation (explicit) Devin: microVM isolation (Devin Environments) All others: no container isolation — tool scope or process-level only

ACP (Agent Client Protocol)

Only Mistral Vibe in this batch implements ACP (agent-client-protocol==0.9.0). It ships both a standard CLI (vibe) and an ACP server (vibe-acp). This is the only framework in the batch enabling agent-to-agent inter-process communication via a standard protocol.

Most Interesting Find

Factory Droid's hooks-as-enforcement philosophy is the most intellectually distinctive position in the batch:

"By encoding these rules as hooks rather than prompting instructions, you turn suggestions into app-level code that executes every time it is expected to run."

This is a falsifiable claim with real architectural implications. It positions prompting as unreliable governance and hooks as reliable governance. No other framework in the corpus makes this argument this explicitly, and Factory has built an entire hook system (9 events) around it.

The secondary interesting find: Zed's recursive skill bootstrap — the built-in create-skill skill that teaches the agent how to create new skills, enabling dynamic skill catalog expansion within a session.

Tier C Items

Antigravity (Google)

• Homepage (antigravity.google/) was unreachable during analysis
• Evidence inferred from companion extension (antigravity-link): CDP port 9000, VS Code fork architecture
• A companion tool (antigravity-link, 180 GitHub stars) reveals the IDE exists and accepts CDP connections
• Insufficient public documentation to write full 11-file analysis

Frontman

• Homepage returned empty/unusable content
• No public documentation, no GitHub repository found
• Likely very early stage or private beta
• Insufficient public documentation to write full 11-file analysis

Cross-References to Other Batches

• superpowers (Batch 1): Factory Droid uses the same SKILL.md convention; Factory is the commercial platform layer on top of what superpowers implements as an open standard
• kiro (Batch N): Both kiro and Factory Droid are commercial closed-source platforms; Factory's distinctive dimension is hooks-as-enforcement + Missions; Kiro's is spec-steering workflow
• Devin vs. Cosine: Both are cloud coding agents targeting autonomous execution; Devin uses standard frontier models; Cosine bets on proprietary Lumen models trained on production outcomes
• Zed vs. Factory Droid: Both use AGENTS.md + SKILL.md; Zed is open-source IDE; Factory is closed commercial platform — the same primitives serve very different product strategies

Evidence Quality Notes

• All open-source frameworks (trae-agent, mistral-vibe, zed, qodo, antigravity-link): full source analysis via gh api
• Closed-source frameworks (devin, cosine, droid-factory): analysis from public marketing sites, docs sites, llms.txt indexes, and published documentation pages
• Factory Droid: particularly rich documentation available at docs.factory.ai including full hook guide, custom droids format, missions documentation, and mixed-models guide
• Cosine: limited public technical detail; product described primarily through marketing copy and a few technical blog posts
• Devin: substantial documentation at docs.devin.ai including skills.md, knowledge.md, AGENTS.md conventions