Part 7 · Integration Blueprints
The first six parts are a reference. This part is a cookbook — seven end-to-end blueprints that weave sandbox, permission, event, and skill into real customer scenarios.
Each blueprint answers the same four questions:
- Who & why — what kind of product / what pain
- Architecture — where Agentao sits, what it talks to
- Key code — the 50–150 lines that matter
- Pitfalls — what tends to break on day two
Key terms in this Part
- In-product assistant — chat / agent embedded inside an existing SaaS UI; the most common shape · §7.1, G.4
- IDE plugin (ACP) — host = editor, agent = subprocess speaking ACP; uses
session/load+request_permission· §7.2, G.3 - Ticket automation — async handler reading from a queue;
prompt_oncestyle, no streaming UI · §7.3, G.4 - Data workbench — interactive analyst session with shell + sandbox + skills · §7.4
- Batch scheduler — cron-driven
prompt_oncefor offline / nightly jobs; no end-user · §7.5, G.4 - WeChat intelligent bot (ilink-style) — long-polling personal-account bot API; one agent per message; contact-scoped permissions · §7.6
- Multi-agent kanban scheduling — external derivative project (
agentao-kanban); board-driven scheduler overplanner/worker/reviewersub-agents; per-card git worktree · §7.7
The seven blueprints
| # | Blueprint | Integration mode | Star extensions |
|---|---|---|---|
| 7.1 | SaaS in-product assistant | In-process SDK + FastAPI | Custom tool + PermissionEngine |
| 7.2 | IDE / editor plugin | ACP stdio | session/load + request_permission |
| 7.3 | Customer-support / ticket automation | In-process SDK | Custom tools hitting CRM |
| 7.4 | Data analyst workbench | In-process SDK | Shell + sandbox + custom skill |
| 7.5 | Offline batch / scheduled jobs | prompt_once | Skills + cron |
| 7.6 | WeChat intelligent bot (ilink-style) | asyncio long-poll daemon | WeChatClient Protocol + contact-scoped permissions |
| 7.7 | Multi-agent kanban scheduling (external repo) | External orchestrator over agentao sub-agents + ACP CLIs | planner/worker/reviewer routing + per-card git worktree |
How to read this part
- If you already picked your scenario, jump straight to that section.
- If you're undecided, 7.1 covers the most common case (in-product assistant) — the other five are specializations.
- Every blueprint links back to the relevant reference sections so you can drill down when needed.
Choose by product shape
| Product you are building | Start with | Why |
|---|---|---|
| Chat assistant inside a SaaS page | 7.1 | Covers the main path: web UI, FastAPI, tools, and permissions |
| IDE, editor, or desktop host | 7.2 | ACP, stdio, permission requests, and session loading are the key differences |
| Queue-driven background automation | 7.3 | Focuses on idempotency, retries, CRM tools, and unattended policy |
| Analyst workbench or notebook-like product | 7.4 | Focuses on shell, sandboxing, file isolation, and analysis skills |
| Nightly jobs, reporting, or offline processing | 7.5 | Focuses on prompt_once, scheduling, budgets, and failure handling |
| IM / WeChat / enterprise messaging bot | 7.6 | Focuses on long polling, contact-scoped permissions, and per-message isolation |
| Multi-agent system over a queue of work (CI, eval, autonomous research) | 7.7 | Focuses on external orchestrators, role routing across sub-agents and ACP CLIs, per-task isolation |
Runnable code
Blueprints 7.1–7.6 ship as self-contained projects inside the main repo at examples/ — each subdirectory (saas-assistant/, ide-plugin-ts/, ticket-automation/, data-workbench/, batch-scheduler/, wechat-bot/) is a standalone uv run / npm run project. Every blueprint page below links to its matching subdirectory.
Blueprint 7.7 lives in a separate repository — jin-bo/agentao-kanban — because it's a derivative project with its own release cadence, not an in-tree example.