Decide the overall shape of an agentic LLM application across five axes: interaction transparency, state/history/memory, context & caching economics, capability integration (tools/skills/MCP/secrets), and execution substrate & side effects. Use when architecting a new coding-agent console or a non-coding agent (research, document, image, or data-artifact producer) before implementation begins, or when auditing whether an existing agentic app hides its reasoning, treats the transcript as its only state, has no context/caching strategy, bolts on tools/secrets with no custody model, or ships side effects with no isolation, no human gate, or no receipt. NOT for implementation stacks once the shape is decided (rust-tauri-development), pure model/router selection (llm-router), prompt wording (prompt-engineer), or multi-agent coordination mechanics once the app-level shape exists (multi-agent-coordination, swarm-invocation-designer).
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Decide the overall shape of an agentic LLM application across five axes: interaction transparency, state/history/memory, context & caching economics, capability integration (tools/skills/MCP/secrets), and execution substrate & side effects. Use when architecting a new coding-agent console or a non-coding agent (research, document, image, or data-artifact producer) before implementation begins, or when auditing whether an existing agentic app hides its reasoning, treats the transcript as its only state, has no context/caching strategy, bolts on tools/secrets with no custody model, or ships side effects with no isolation, no human gate, or no receipt. NOT for implementation stacks once the shape is decided (rust-tauri-development), pure model/router selection (llm-router), prompt wording (prompt-engineer), or multi-agent coordination mechanics once the app-level shape exists (multi-agent-coordination, swarm-invocation-designer).
license
Apache-2.0
allowed-tools
Read,Write,Edit,Bash,Grep,Glob
metadata
{"category":"Agent & Orchestration","tags":["agent-architecture","transparency","episodic-memory","capability-integration","execution-substrate"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"agentic-coding-ux-designer","reason":"Turns the transparency axis into a concrete flow (plan/apply/review, checkpoint, receipt UI)."},{"skill":"agent-work-receipt-designer","reason":"Defines the receipt schema this skill requires at the execution-substrate finish line."},{"skill":"human-gate-designer","reason":"Designs the actual approval UX for the side-effect human checkpoint this skill mandates."},{"skill":"episodic-memory-algorithms","reason":"Supplies the TTL/blob-promotion/recall mechanics behind the state/memory axis."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]","[Truncated]"],"produces":["[Truncated]","[Truncated]"]}}
Agentic App Architecture
Decide the shape of an agentic app — what it shows, what it remembers, what it costs, what powers it has, and where its actions land — before writing implementation code.
Use This For
Architecting a new coding-agent console (Claude-Code-like), operator surface, or pd-console-style fleet UI from scratch.
Architecting a non-coding agent that produces documents, images, data artifacts, or self-authored tools at runtime.
Deciding memory model, context/caching strategy, and MCP/tool/secret wiring before implementation locks them in.
Auditing an existing agentic app for a hidden-reasoning, transcript-only-state, unbounded-context, or ungated-side-effect design flaw.
Reconciling a coding agent's git/worktree substrate against a non-coding agent's artifact-and-receipt substrate in the same product.
Do Not Use This For
Picking a UI framework, desktop shell, or deployment stack once the architecture axes are decided (rust-tauri-development).
Choosing which model or router serves a request (llm-router) or wordsmithing the system prompt (prompt-engineer).
Designing the moment-to-moment protocol between cooperating agents once the app-level shape already exists (multi-agent-coordination, swarm-invocation-designer).
Five-Axis Decision Flow
flowchart TD
A[Interaction surface & transparency] --> B[State, history & memory]
B --> C[Context & caching economics]
C --> D[Capability integration: tools/skills/MCP/secrets]
D --> E[Execution substrate & side effects]
E --> F{Coding agent?}
F -->|yes| G[Worktree isolation + PR finish line]
F -->|no| H[Artifact + self-authored-tool substrate]
G --> I[Ship with receipts]
H --> I
Choose the interaction model. Decide what surfaces the agent's thinking and tool calls (inline, streaming, collapsible workbench), whether a plan is shown before acting, and how a human interrupts mid-run. A hidden-hands agent is untrustable by construction — see references/interaction-surface-and-transparency.md.
Choose the state/memory model. Decide whether history is durable, whether threads can fork, and whether salient facts get promoted to episodic memory with TTLs. The transcript is not the whole state — see references/state-memory-and-context.md.
Choose the context/caching strategy. Decide the caching approach (the Anthropic prompt cache has a ~5-minute TTL, so poll/sleep cadence and context stability matter), the eviction/summarization policy, and what gets promoted out of the window into memory.
Wire capabilities with custody. Decide tool schemas (lazy-load if the toolset is large), skill packs, an MCP topology (small always-on global core, per-project specialists — see references/capabilities-and-execution-substrate.md), and how secrets reach tool calls without landing in argv, logs, or the transcript.
Choose the execution substrate and side-effect gates. For coding agents: one worktree per writer, advisory claims, a PR finish line with artifact-backed validation, never force-push or touch main. For non-coding agents: durable artifacts (documents/images/data/rendered web artifacts) or even agent-authored tools, each gated by a human checkpoint for irreversible/outward-facing actions.
Ship with receipts. Every side-effecting run leaves a durable, artifact-backed receipt — not a chat message claiming success.
Audit before you build, and again before you ship. Run scripts/agentic_app_audit.mjs against the spec at design time, and again against the shipped reality.
Output Contract
Produce an architecture decision covering all five axes plus a machine-checkable audit:
Use scripts/agentic_app_audit.mjs to score a spec matching schemas/agentic-app-spec.schema.json and return { pass, coverageByAxis, findings, recommendations }.
Anti-Patterns
Chat Box With Secret Hands
Novice: "Put an AI chat next to the work surface and let it act; the transcript is enough of a UI."
Expert: Hidden thinking and hidden tool calls make an agent both untrustable and un-steerable. Reasoning and tool use must be surfaced inline or in an inspectable pane, with a plan shown before consequential action.
Detection: agentic_app_audit.mjs fires hidden-thinking-or-tool-use (critical) whenever thinkingVisible or toolUseVisible is false.
Transcript Is The Whole State
Novice: "The conversation history is the app's memory; there's nothing else to design."
Expert: Durable history, thread forking to explore alternates, and episodic memory (TTL'd, relevance-recallable, promoted out of the raw transcript) are separate, necessary pieces of state design — not features of a good chat UI.
Detection: agentic_app_audit.mjs fires transcript-only-state (critical) whenever both forking and episodicMemory are false.
Side Effects With No Gate, No Isolation, No Receipt
Novice: "The agent can write files, call APIs, and merge its own work; we'll review if something looks wrong."
Expert: Every side-effecting action needs isolation (a worktree for coding agents, a sandbox for non-coding agents), a human checkpoint before anything irreversible or outward-facing, and a durable artifact-backed receipt afterward — especially for coding agents writing to a shared checkout.
Detection: agentic_app_audit.mjs fires no-execution-isolation, no-human-gate-on-side-effects, or no-artifact-receipt (critical for agentType: coding, high otherwise) whenever isolation, sideEffectHumanGate, or artifactReceipts is false.
examples/expected-output.md — Example Output: Agentic App Architecture — Scenario: "Harbor Scout" — a non-coding Port Daddy fleet agent that researches a competitor's release notes, writes a summary report artifac
references/capabilities-and-execution-substrate.md — Capabilities & Execution Substrate — Use this when wiring an agent's powers (tools, skills, MCP servers, secrets) and when deciding where its actions actually land.
references/interaction-surface-and-transparency.md — Interaction Surface & Transparency — Use this when deciding what an agentic app shows the human — thinking, tool use, plans, and steering — before it renders a single screen.
references/state-memory-and-context.md — State, Memory & Context — Use this when deciding what persists beyond the current turn, what gets promoted to durable memory, and how the context window and prompt ca
templates/output-template.md — Agentic App Architecture Decision — [One-sentence description of the app being architected, and whether it is a coding agent or non-coding agent.] - Transparency: [what the
