Before you can understand the field, you have to understand why it's confusing.
The same names appear everywhere — "Conductor", "Blueprint", "Ralph" — attached to completely unrelated projects.
These six things are all called "Conductor". None of them knows the others exist.
The confusion isn't limited to "Conductor". The corpus has six things called "Conductor",
eight things called "oh-my-something", two unrelated "Blueprint" projects, and three independent "Playbook" documents.
The naming collision is the field's surface problem. The structural problem goes deeper.
Screen 2 of 5
21 archetypes. Most people only know three.
When people talk about "AI coding frameworks", they're usually thinking about three things:
skills/commands packs, multi-agent orchestrators, and spec-driven workflows.
The atlas identified 21 distinct archetypes — six of them didn't exist in the Phase A seed taxonomy.
The newcomers include: cross-vendor model routers, compiled enforcement frameworks,
self-evolving plugins, token-efficient encoders, OS-distro substrates, and LLM-evaluated hooks.
Five races. Three of them did not exist last year.
Phase A (11 seeds) framed this space as a spec-driven-development problem.
Phase B (319 frameworks) uncovered an operational layer the seeds had no vocabulary for —
orchestrators, PR-lifecycle daemons, governance frameworks, UI passthroughs.
Phase D (323 more) added three new races that didn't exist in Phase A's taxonomy:
cross-vendor model routing (20 frameworks), token-efficient symbolic encoding (14 frameworks, led by rtk-token-killer at 54k★),
and LLM-evaluated hooks (4–5 frameworks, now Archetype A21).
Phase A
11
Seeds
The 11 canonical reference frameworks. 5 archetypes. Spec-driven framing.
32 more batches. Three new races surfaced. 6 new archetypes. 21 total.
Phase E
Synthesis
Cross-corpus synthesis. The 5-framework reference baseline. Top-15 deep-dive list.
Screen 4 of 5
The most popular framework solves one trick.
GitHub stars in this corpus correlate with SEO reach and social sharing — not engineering rigor.
The five highest-starred Phase D entries are aggregation packs and UI dashboards.
rtk-token-killer at 54k★ is the most-starred Phase D framework —
and it does exactly one thing: intercept Bash output and compress it.
Meanwhile, goopspec (29★) has 16 agents, 1,660 passing tests, and cost-tier executor mapping.
The /atlas scatter makes this disconnect visible.
Five frameworks form a defensible reference stack.
The Phase E synthesis forced a choice: if you had to merge five frameworks into a flawless reference design,
which would you pick? Each one solves exactly one dimension no other framework covers as well.
Together, they cover all 25 rubric dimensions.
Phase D adds two more — Trellis and gh-aw-githubnext — for a 7-framework defensible baseline.