Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More
77 contributions to AI Automation Society
Make an OpenAI agent run a task every morning with one API request
Want an agent to send you a status summary every morning, or follow up on something later, without n8n or an external scheduler? The OpenAI Agents API now has scheduled automations: you create one with a single request and OpenAI starts the agent session for you. It is in public beta, added on Oct 23 per the changelog. Here is how to set one up (docs: https://developers.openai.com/api/docs/guides/agents-api/sessions/automations ): 1. Create an agent with the instructions and tools the task needs, and note its agent_id. 2. Pick a target. new_session starts fresh and needs agent_id, environment and your message in target.input. existing_session keeps earlier context and needs session_id plus reminder text in target.instructions. 3. Pick a schedule. periodic takes a cron expression (a five-field pattern for when to run) plus a time zone, for example "0 10 * * *" with "America/New_York" means every day at 10 AM. once takes run_at, a Unix timestamp in seconds, set in the future. 4. Send POST /v1/agents/automations with the header OpenAI-Beta: agents=v1 (the SDKs add it for you). The doc's daily example body: 5. {"type":"periodic","name":"Daily session","expression":"0 10 * * *","timezone":"America/New_York","target":{"type":"new_session","agent_id":"agt_123","environment":{"type":"openai_hosted"},"input":[{"role":"user","content":[{"type":"input_text","text":"Summarize the current status and next steps."}]}]}} 6. Test it. The automation is saved as active but does not run right away, so trigger a manual run with POST /v1/agents/automations/{id}/invocations (no body). Check history with GET on the same path. Pause, resume and delete have their own endpoints. Gotchas: a run with status completed only means the session accepted the task, not that the agent finished. Recurring runs get a small random delay, and if the session is busy the task waits. A scheduled run also does not supply a missing approval, so check the approvals page: https://developers.openai.com/api/docs/guides/agents-api/sessions/approvals
Let an AI agent answer from your own PDFs, with page citations, in 5 steps
Want an agent that answers from your own documents but don't want to run a vector database (a database that stores text as numbers so it can be searched by meaning)? Gemini API File Search does the chunking, indexing and retrieval for you, and the answers come with sources. Cost: storage is free and query-time embeddings are free. You pay only when files are indexed: gemini-embedding-2 is $0.20 per 1M tokens on the paid tier ($0.10 in Batch), and text input is free on the free tier (https://aistudio.google.com/docs/pricing). 1. Get a key and install the SDK. In Google AI Studio, open the API keys page and click Create API key. Then run pip install -U google-genai and export GEMINI_API_KEY="YOUR_API_KEY" (JS: npm install @google/genai). https://aistudio.google.com/docs/quickstart 2. Create a File Search store (the place your documents get indexed): client.file_search_stores.create(config={'display_name': 'your-fileSearchStore-name'}). The default embedding is text-only. Add 'embedding_model': 'models/gemini-embedding-2' if you also want PNG/JPEG images searchable. REST: POST https://generativelanguage.googleapis.com/v1beta/fileSearchStores 3. Upload your PDFs with client.file_search_stores.upload_to_file_search_store(file=..., file_search_store_name=..., config={'display_name': ...}). It returns an operation, so poll client.operations.get(operation) until operation.done. Optional: add custom_metadata at import, like {"key": "author", "string_value": "Robert Graves"}. 4. Ask a question with client.interactions.create(model=..., input=..., tools=[{"type": "file_search", "file_search_store_names": [store.name]}]). To search only some documents, pass metadata_filter, for example 'author="Robert Graves"'. 5. Read the citations in the annotations of the text content blocks. Each file_citation can include file_name, source, document_uri, page_number (for PDFs) and custom_metadata. Show them to your users as sources.
OpenAI's own tests show its text watermark weakens after word swaps
Article 50 of the EU AI Act obliges providers of AI systems that generate synthetic audio, image, video or text to mark outputs in a machine-readable format, so they can be detected as artificially generated. The same paragraph asks that the solutions be effective, interoperable, robust and reliable, but only as far as technically feasible. The text says: “Providers shall ensure their technical solutions are effective, interoperable, robust and reliable”, followed by the qualifier “as far as this is technically feasible”. The page sets no numeric accuracy or robustness threshold. You can read the full article at my news agentic website here: The Frame News
Make your Zapier AI automation wait for your OK in 5 steps
Your Zap lets AI draft an email or a reply, and then it sends it with nobody looking. Zapier has a built-in step called Human in the Loop that pauses the Zap until a person approves, declines or edits the data. It takes about 3 minutes to add. What you need: a paid Zapier plan. Human in the Loop is not on Free; it is included in Professional, Team and Enterprise. 1. Add a Human in the Loop step and choose Request Approval. Put it before the risky action, for example before the step that sends the AI-drafted email. 2. Configure the request. Write the approval request message, add the content to review (map the AI draft from the earlier step), set the Approve and Decline button labels, choose "Action if reviewer declines" (Continue run or Stop run) and "Let Reviewer edit content?" (Yes or No). 3. Choose how reviewers are notified: Email, Slack, or Trigger a Zap. Trigger a Zap needs a second Zap that starts with the New Approval Requested trigger. 4. Set the reviewers, the timeout and an optional reminder. Only one selected reviewer needs to respond. "On timeout" is either Skip and continue or End run, and the timeout can be set in minutes, hours, days or weeks (minimum 1). 5. Test the step. Testing creates sample fields and values but does not notify reviewers. Use the preview to see a sample notification, or run the whole Zap end to end to try a real approval. Docs: https://help.zapier.com/hc/en-us/articles/38731463206029/ Gotchas from the docs: reviewers need a Zapier account (a free one is enough) and the Zap must be shared with them. On a Pro account you can only send requests to yourself. The step cannot go inside or after a Looping step. What is the first Zap where you would add an approval step? I'm building an AI agency, https://ethiaagency.com, and sharing here what has been most useful to me.
Start a Claude Code cloud task from any automation with one HTTP POST
Your laptop has to be open for Claude Code to work on a task. A scheduled task (called a routine) runs on Anthropic's cloud instead, and you can start it from anywhere that can send an authenticated HTTP request. That's 6 steps. Heads up: it is a research preview, so behavior, limits and the API may change. You need a paid claude.ai plan (Pro, Max, Ultra, Team or Enterprise) and a GitHub repo Claude can clone. Runs use your subscription usage. Docs: https://code.claude.com/docs/en/routines 1. Go to claude.ai/code/routines and create a new task. Write a self-contained prompt: it runs on its own, so say exactly what to do and what success looks like. 2. In Advanced settings, pick your GitHub repo(s) and a cloud environment, and remove connectors (the links to your other apps) the task does not need. 3. In the prompt, point at the incoming text, for example: "Investigate the alert described in the routine-fire-payload block". Without that line, Claude treats your text as background only. 4. Open the task, click the menu next to its name, then Edit. In the Trigger row click +, choose API, copy the Fire URL and click Generate token. The token is shown once, so store both as secrets in your automation tool. This step only works on the web, not from the CLI. 5. Send a POST from your automation (API reference: https://platform.claude.com/docs/en/api/claude-code/routines-fire): curl -X POST https://api.anthropic.com/v1/claude_code/routines/$ROUTINE_ID/fire -H "Authorization: Bearer $ROUTINE_TOKEN" -H "anthropic-version: 2026-10-21" -H "Content-Type: application/json" -d '{"text": "Sentry alert SEN-4521 fired in prod. Stack trace attached."}' The text field is optional, freeform, up to 65,536 characters. The anthropic-beta header shown on the routines page is optional.
0 likes • 12d
@Kunal Goyal Good call. The plan is one small table per routine: event id, the session URL from the fire response, and a status. Anything still pending after a set time gets flagged, so a quiet run never goes unnoticed.
0 likes • 12d
@Wasilis Paliakoudis Agreed, that's the gotcha that worries me most. I gate on the event id and mark it pending before the POST, so a retry mid-flight gets skipped, plus a hard daily ceiling. One edge case worth adding: a timed-out POST may still have created the session, so blind retries are exactly how the duplicate happens.
1-10 of 77
Theo Ethia
6
1,134 points to level up
@ethia-x-8995
Im Theo, from ETHIA Agency. I develop software and businesses

Active 2d ago
Joined Jun 2, 2026
Powered by