| name | agent-native-architecture |
| description | Build applications where agents are first-class citizens. Use this skill when designing autonomous agents, creating MCP tools, implementing self-modifying systems, or building apps where features are outcomes achieved by agents operating in a loop. |
Purpose
Internal architecture reference/helper for agent-native app design: parity, primitive tools, prompt-native features, shared workspaces, production guardrails, and improvement over time. For full-codebase agent-native architecture audits, read references/audit-playbook.md as the audit adapter.
Canonical Taxonomy
Canonical taxonomy for adjacent audit/review/research assets: Action parity, Primitive tools, Shared workspace, Context injection, Prompt-native features, Production guardrails, Eval readiness. Local labels may vary, but must map back here.
Invocation Boundary
agent-native-architecture is an internal architecture reference/helper, not a public spec-* workflow. Upstream workflows may read it as bounded context; do not imply a standalone public command.
When To Use
- Agent-native design/review, primitive tools, MCP surfaces, dynamic capability discovery.
- Action parity, self-modification, prompt-native, shared-workspace, or production-readiness questions.
When Not To Use
- Public entrypoint, replacement for
spec-plan/spec-work/spec-code-review/spec-doc-review, provider-specific contract fields, or generated runtime mirror edits.
Inputs
Bounded architecture question plus relevant source files, product surfaces, tool definitions, prompts, tests, workflow artifacts, and source-scoped advisory external facts.
Outputs
Source-backed architecture guidance, checklists, capability maps, reference pointers, limitations, and required handoff to public spec-* workflows when this reference is not enough.
Workflow
- Confirm bounded context or ask one short clarification.
- Read only the matching reference track.
- Apply the canonical taxonomy and relevant guardrails.
- Return source refs, limitations, and public workflow handoff when needed.
Failure Modes
Generic advice from missing context; unsafe autonomy from weak approvals; provider overfit; invisible prompt drift from missing evals; generated-runtime edits instead of source.
Reference Routing
Read only the track that matches the task:
- Overview:
references/agent-native-principles.md
- Architecture/shared workspace:
references/architecture-patterns.md, references/files-universal-interface.md, references/shared-workspace-architecture.md, references/checklists.md
- Tools/MCP/domain tools:
references/mcp-tool-design.md, references/from-primitives-to-domain-tools.md
- Prompts/context/loop:
references/system-prompt-design.md, references/dynamic-context-injection.md, references/agent-execution-patterns.md
- Action parity/full app audit:
references/action-parity-discipline.md, references/audit-playbook.md
- Production autonomy or self-modification:
references/runtime-production-guardrails.md, references/self-modification.md, references/product-implications.md
- Mobile/testing/refactoring:
references/mobile-patterns.md, references/agent-native-testing.md, references/refactoring-to-prompt-native.md
Route and boundary examples live in evals/examples.json; use them for review/eval readiness, not as runtime state.
Runtime/Source Boundary
The source of truth is skills/agent-native-architecture/SKILL.md, references/, evals/examples.json, and centralized governance in src/cli/contracts/dual-host-governance/skills-governance.json. Do not add a per-skill manifest as a second governance source. Generated runtime mirrors are not source-of-truth and must not be hand-edited as source fixes.
