Agent Skills

Preview. Agent Skills is a NAMS-backend feature in preview. The backend contract (REST endpoints and MCP tools) is stable enough to build against, but the dashboard experience is still evolving and the surface is not yet exposed in the Python or TypeScript SDKs. There is no equivalent on the bolt backend.

An Agent Skill turns accumulated agent memory into a small, trustworthy, reusable procedure another agent can load. Where the three memory layers capture what happened in one workspace, a skill is the distilled how-to — a portable package other agents (or the same agent in a fresh context) can pick up without starting cold.

The problem skills solve

A NAMS workspace accumulates three kinds of memory: short-term messages, long-term POLE+O entities, and reasoning traces (steps and tool calls). That experience is valuable but trapped — it lives in one workspace’s graph, in a form no other agent can consume directly. Skill distillation is the transform that lifts a scoped slice of that graph into a portable, spec-compliant Agent Skill (a SKILL.md package following the agentskills.io format) that any agent can load.

Create → review → use

The feature is organised around three jobs:

Create: Distil a skill from a scope — the whole workspace, a single entity, or an ontology class. NAMS snapshots the relevant memory, synthesises a procedure grounded in it, runs quality gates, and packages the result.
Review: A human approves or rejects a distilled skill before it is published, so the published library stays healthy. Skills move through a lifecycle: draft → in_review → published (with rejected as a terminal state).
Use: Download the published SKILL.md package (a ZIP), or discover what exists over MCP, and load it into an agent.

Fidelity via provenance

The defining property of a distilled skill is that the graph is the authority and the LLM is only a writer working from cited sources. Every claim and every procedure step is GROUNDED_IN specific source node ids — messages, entities, or reasoning steps in the workspace. That makes a skill auditable, not invented: you can trace any statement back to the memory it came from (explain-provenance), and a published package ships a provenance.json mapping. Distillation enforces quality gates on this grounding — a run whose grounding or coverage is too low is withheld (often with a suggestion to split it into smaller procedures) rather than published.

How a skill relates to memory

Skills are a derived projection of the existing memory model, not a new memory layer:

Concept Relationship to skills

Concept	Relationship to skills
Short-term / long-term / reasoning memory	The raw material. A skill’s claims and steps are grounded in `:Message`, `:Entity`, and reasoning (`:AgentStep` / `:ToolCall`) nodes.
Reasoning traces	The primary source for procedure steps — a recorded tool call becomes a typed script step; a reasoning step becomes a judgement step.
Ontology	Can be used as a distillation scope (`scope_type = ontology_class`), so you can distil "everything we know about this class of thing".
New graph objects	Distillation adds `:Skill` → `:SkillVersion` → `:SkillStep` / component nodes plus a `:DistillationRun`, co-located with the memory they cite.

Short-term / long-term / reasoning memory

The raw material. A skill’s claims and steps are grounded in :Message, :Entity, and reasoning (:AgentStep / :ToolCall) nodes.

Reasoning traces

The primary source for procedure steps — a recorded tool call becomes a typed script step; a reasoning step becomes a judgement step.

Ontology

Can be used as a distillation scope (scope_type = ontology_class), so you can distil "everything we know about this class of thing".

New graph objects

Distillation adds :Skill → :SkillVersion → :SkillStep / component nodes plus a :DistillationRun, co-located with the memory they cite.

What ships in a skill package

A downloaded skill bundle contains a SKILL.md (YAML frontmatter carrying the grounding score and provenance id), a typed procedure (a step graph, or prose for simpler skills), provenance.json, and reference docs (domain model, exemplars, procedures). Published versions are attested — signed with a detached JWS you can verify offline against a public JWKS — so a consumer can confirm a package’s integrity before loading it.

Maturity & limits

Distillation is on by default.
Composition (extracting a sub-procedure that other skills CALLS) is off by default.
Execution is dry-run planning only and gated off by default — NAMS will order and validate a skill’s steps but does not invoke tools. Running the procedure is the loading agent’s responsibility.
Whole-workspace distils tend to be withheld once a workspace holds more than one distinct procedure — prefer a narrower scope.

See Skills API for the REST + MCP surface and the Skills Quickstart for an end-to-end walkthrough.