技術OpenAI Cookbook Agents 主題 2026-09-10 新增三篇之一AI自動化工作流

OpenAI Cookbook：用 Agents API 做一個會查資料、開 PR 的 Slack 機器人

Build a Slack bot with the Agents API

來源：OpenAI Cookbook・發布 2026-09-10・本站摘要分析（非全文轉載）

📌 一句話啟示：把 AI agent 接進公司已在用的 Slack，是上班族最容易說服主管的導入切入點。

摘要分析

OpenAI Cookbook 2026 年 9 月新增的 Agents API 應用範例。同事在 Slack 標記機器人就能搜尋對話、調查專案、分析資料或準備 GitHub PR；每條 thread 對應一個 Agents API session 與獨立沙箱，追問會延續同一調查。步驟：複製 .env 範例 → docker build 沙箱映像 → 填 API key 與受限的 executor key → 用 manifest 建 Slack app 取得 token → uv run main.py 以 Socket Mode 啟動，不需公開 webhook。可選接 Notion、Google Drive、GitHub。

需 Python 3.14+、uv、Docker；cp .env.example .env
docker build -t agent-api-sandbox:latest examples/agents_api/sandboxes/docker/application_managed
填 OPENAI_API_KEY 與 OPENAI_EXECUTOR_API_KEY（只有後者進沙箱）
用 manifest 建 Slack app，設 SLACK_BOT_TOKEN、SLACK_APP_TOKEN
uv run examples/agents_api/apps/slack_bot/main.py，thread 內 @機器人即可用

One of the Agents API application recipes OpenAI added in September 2026 (alongside an SRE incident-response agent and a data analyst). Teammates tag the bot in Slack to search conversations, investigate projects, analyze data or prepare a GitHub pull request; each Slack thread maps to its own Agents API session and isolated sandbox workspace, so follow-ups continue the same investigation. Slack supplies the trigger and interface, the Agents API keeps conversation state, runs the agent and streams progress, and your application controls Slack access, credentials and sandbox lifecycle. Steps: copy the .env example, build the Docker sandbox image, set OPENAI_API_KEY plus a restricted executor key with api.agents.environments.connect permission, create the Slack app from the provided manifest and collect the bot and app-level tokens, then start with uv run main.py in Socket Mode (no public webhook needed). Optional shared tools include Notion, Google Drive and GitHub.