Explore skill-graph concept as evolution of bundles (atoms / molecules / compounds)

[luongnv89]

Type

feature (medium confidence) — idea / reference, not yet a concrete spec

Description

Capture the skill-graph concept as a potential future direction for ASM bundles. No clear implementation vision yet — filing this to keep the idea searchable and linked to the existing bundle work (#202, #238) so we can return to it later.

The idea (summary)

Shiv Sakhuja's framing (source tweet) proposes that skills compose best when organized into three tiers rather than a flat dependency graph:

• Atoms (capabilities) — narrow, single-purpose, near-deterministic building blocks. Don't call other skills.
• Molecules (composites) — orchestrate 2–10 atoms with explicit chaining. Most of the composition logic lives in the skill, minimizing runtime judgment.
• Compounds (playbooks) — higher-level orchestrators that run multiple molecules. Meaningful agent autonomy, human-driven.

The key insight: naive deep skill-graphs become unreliable past a few hops because agents don't consistently traverse long dependency chains. The tiered model keeps graph depth shallow within a tier and makes composition explicit upward.

Why this is interesting for ASM

Today's bundle (#202) is a flat set of skills installed together. A skill-graph bundle would be a composed workflow:

Imagine selecting a set of existing skills, then re-writing them so you have a bundle (skill-graph) that automates some specific workflow.

Instead of "install these 5 skills," a skill-graph bundle would ship:

• The selected skills as atoms (normalized / wrapped, not mutated)
• A generated molecule that explicitly chains them for a named workflow
• Optional compound that wraps molecules for human-driven orchestration
• A graph manifest describing tiers, nodes, and edges

Possible surfaces (not committed — just brainstorming)

• Extended bundle schema: add `tier` and `depends_on` per-skill in the bundle JSON
• Website: visualize a bundle's graph (tiered nodes, edges)
• CLI: `asm bundle compose` that takes selected atoms and scaffolds a molecule/compound wrapper
• Skill-authoring UX: optional `tier` hint in SKILL.md frontmatter
• Validator: atoms don't call other skills; molecules explicitly name their atoms; depth ≤ ~3

Open questions

• Do atoms get rewritten or just wrapped? (Wrapping is safer — preserves upstream skills verbatim.)
• Is the unit of distribution one bundle manifest referencing external skills, or a folder containing generated molecule/compound skills alongside atom copies?
• How does this interact with the existing bundle schema from Add skill bundle management (create, edit, install, export) #202?
• Is the tiering enforced or advisory?
• Do we need evaluation harness support to measure bundle reliability (since the post flags that testing composed skills is non-trivial)?

Status

Reference only. Not prioritized, no owner, no concrete spec. Revisit after #202 and #238 land and we have real usage data for flat bundles.

Reporter Context

adopting skill-graph idea https://x.com/shivsakhuja/status/2047124337191444844

[On the concept]: I am building asm (agent skill management), I have just introduced skill bundle (install a set of skills with 1 command), but this skill-graph is very interesting idea, can be a new direction for skill bundle. Imagine you can select a set of existing skills, then re-write them so that you can have a bundle (skill-graph) to automate some specific workflow.

[On scope]: just keep it [as a] reference, I don't have a clear vision for that right now, but could be an interesting concept to be adopted into asm.

Related Issues

• Add skill bundle management (create, edit, install, export) #202 — Add skill bundle management (create, edit, install, export)
• Add website UI to build custom skill bundles (cart flow) #238 — Add website UI to build custom skill bundles (cart flow)

Acceptance Criteria

• Decide whether to extend bundle schema with tier/dependency fields, or keep skill-graph as a separate bundle variant. (low — needs design)
• Prototype one end-to-end example: pick 3–5 atoms, produce a molecule that chains them, validate it runs reliably. (low — needs design)
• Document tier definitions (atom / molecule / compound) in ASM docs with concrete examples drawn from existing skills in the catalog. (medium)
• Decide whether tier is enforced by a validator or advisory frontmatter. (low — needs design)

Metadata

• Priority: low (parked idea)
• Estimated effort: unknown — depends on scope chosen when revisited
• Suggested labels: enhancement, feature, bundles, discussion, rfc