agent-native-architecture

## What It Does Codifies the five principles that distinguish "agent-native" architecture from "agent-as-afterthought" — the patterns the `yellow-review:review:agent-native-reviewer` agent checks against. This skill is the canonical reference; the reviewer agent applies it. ## When to Use - Authoring or modifying agent tool definitions, system prompt construction, or LLM-integration scaffolding. - Designing a feature where users and agents will both act on shared data. - Evaluating whether a UI action has a corresponding agent capability. - Deciding whether a tool should be a primitive or a workflow. ## Usage ### The Five Principles 1. **Action Parity** — Every UI action has an equivalent agent tool. If a user can rename a project via a button, the agent must have a `rename_project` tool. Exceptions are intentional human-only flows (CAPTCHA, 2FA, OAuth consent, biometric auth) and purely cosmetic UI (animations, theme toggling). 2. **Context Parity** — Agents see the same data users see. The system prompt must include available resources (files, entities, recent activity), capability mappings (which tool does what), and domain vocabulary. Static system prompts that don't reflect runtime state produce agents that don't know what exists. 3. **Shared Workspace** — Agents and users operate in the same data space. Agent file operations use the same paths as the UI; the UI observes changes the agent makes (via shared store, file watching, or reactive binding). Anti-pattern: agent writes to `agent_output/` while user works in `documents/` — separate sandboxes break collaboration. 4. **Primitives over Workflows** — Tools are composable primitives whose inputs are data, not decisions. A `store_item(key, value)` tool is correct; a `process_feedback(message)` tool that internally categorizes + prioritizes + notifies is a workflow tool that strips the agent of reasoning agency. **Exception:** Workflow tools are acceptable when they wrap safety-critical atomic sequences (a payment must charge + record + receipt as one unit) or external orchestration the agent shouldn't step-through (a deploy tool). 5. **Dynamic Context Injection** — System prompts include runtime app state, not just static instructions. The prompt at build-time names the agent's role; the prompt at request-time names the user's current project, recent files, available entities, and which capabilities apply to this specific call. ### The Noun Test For every domain noun in the app (project, document, task, message, report — whatever the entities are), verify the agent can: 1. **Know what it is** — domain vocabulary in the system prompt 2. **Interact with it** — at least one tool whose target is this noun 3. **Discover the capability** — the tool is named in the prompt or surfaced in onboarding A noun that fails all three is a critical gap for must-have entities. ### Anti-Pattern Catalog | Anti-Pattern | Signal | Fix | |---|---|---| | Orphan Feature | UI action with no agent tool | Add tool; document in prompt | | Context Starvation | Agent unaware of resources or terms | Inject resources + vocabulary into prompt | | Sandbox Isolation | Agent reads/writes separate data space | Use shared workspace | | Silent Action | Agent mutates state, UI doesn't update | Reactive binding or file watching | | Capability Hiding | Users can't discover agent capabilities | Surface in agent responses | | Workflow Tool | Tool encodes business logic | Extract primitives; orchestrate in prompt | | Decision Input | Tool accepts decision enum | Accept data; let agent decide | ### Priority Tiers Not every gap is equal — prioritize findings by impact: - **Must have parity:** Core domain CRUD, primary user workflows, actions that modify user data. - **Should have parity:** Secondary features, read-only views with filtering/sorting. - **Low priority:** Settings/preferences UI, onboarding wizards, admin panels, purely cosmetic actions. Critical findings only apply to must-have and should-have tiers; low-priority gaps are observations. ### Stack-Specific Tool Locations | Stack | Agent tools live at | |---|---| | Vercel AI SDK (Next.js) | `tool()` in route handlers; `tools` param in `streamText`/`generateText` | | LangChain / LangGraph | `@tool` decorators; `StructuredTool` subclasses | | OpenAI Assistants | `tools` array in assistant config | | Claude Code plugins | `agents/*.md`, `skills/*/SKILL.md`, tool lists in frontmatter | | Rails + MCP | `tool()` in MCP server definitions; `.mcp.json` | | Generic | `tool(`, `function_call`, `tools:`, tool registration patterns | ### What This Skill Doesn't Cover - **CLI agent-readiness** — see `yellow-review:review:cli-readiness-reviewer` and `yellow-review:review:agent-cli-readiness-reviewer` for CLI-specific agent optimization (non-interactive defaults, structured output, etc.). - **Agent authoring conventions** — see `yellow-core:create-agent-skills` for how to write Claude Code agents and skills. - **MCP integration patterns** — see `yellow-core:mcp-integration-patterns` for ruvector and morph integration patterns.