How to trigger AI agents from webhooks, Slack, and cron

An agent that only answers when someone types into a chat box is a tool. An agent that starts on its own, when a deploy fails, a customer emails, or Monday morning arrives, is an automation. The trigger decides how quickly the agent responds, how much noise it creates, and what can go wrong.

The table compares the main trigger types. The sections below cover how to set up each one, using the senders' own documentation, and what type.com handles for you.

Speed comes first, because senders don't wait for an agent. GitHub expects a 2xx within 10 seconds and treats a slower response as a failed delivery. Stripe says to return a 2xx before any complex logic. An agent run takes much longer than that, so accept the event, queue it, and let the agent run separately.

Then prove the event is real. Most senders sign the raw request body with HMAC-SHA256 and a secret you share: GitHub sends X-Hub-Signature-256, Slack sends X-Slack-Signature with a timestamp, and Stripe sends Stripe-Signature. Compute the signature over the raw bytes before parsing JSON, compare in constant time, and reject old timestamps to stop replays. Stripe's libraries allow five minutes by default.

Finally, expect duplicates. Stripe says endpoints might occasionally receive the same event more than once, GitHub keeps the same X-GitHub-Delivery ID when you redeliver, and the Standard Webhooks spec, which OpenAI and Anthropic follow, keeps the same webhook-id across retries and recommends it as an idempotency key. Store the IDs you've handled and skip repeats, so one alert never starts two agent runs.

On type.com, a webhook automation gives you the URL. Save the sender's shared secret and type.com verifies each request: it accepts HMAC-SHA256 or HMAC-SHA1 signatures of the body and detects the header from the first valid signed request. Senders that sign a timestamp along with the body, such as Stripe, may need a compatible custom sender. Without a secret, treat the URL itself as a credential. A 200 response means type.com accepted the event, not that the agent finished, so check the thread the event created. For engineers, code automations, in early access, can receive webhooks through an adapter that registers the webhook and verifies every request.

Slack's Events API expects an HTTP 2xx within three seconds. If it doesn't get one, it retries up to three times, nearly immediately, after a minute, and after five minutes, with x-slack-retry-num and x-slack-retry-reason headers, so a slow handler can start the same agent run three times. If more than 95% of deliveries fail within an hour, Slack temporarily disables your event subscriptions, unless your app receives fewer than 1,000 events an hour.

Choose the event carefully. app_mention fires only when someone mentions your app in a conversation it belongs to, and Slack itself recommends it if you only want messages sent to your app. The message events fire for every post in a channel, which is rarely what an agent should answer. Verify each request with your app's signing secret: the signature covers v0, the timestamp, and the raw body. If you can't expose a public URL, Socket Mode delivers Events API traffic without one, though Socket Mode apps aren't allowed in the public Slack Marketplace.

In type.com, you map a Slack channel to a Space and choose when the managed agent responds: on @mentions, on new channel messages, or on every message including thread replies. People can stop a response by adding a 🛑 reaction. Code automations can also run on Slack events such as a message containing a keyword or an added reaction, for example starting a job when someone reacts with a check mark.

Every scheduler has its own fine print. A GitHub Actions schedule runs at most every five minutes, in UTC unless you set an IANA time zone, only from the default branch, and it can be delayed or even dropped when GitHub is busy, especially at the top of the hour. Claude Managed Agents scheduled deployments, in beta, take a cron expression and a time zone at minute granularity, add jitter of up to 15% of the interval, and don't backfill triggers missed while paused. Microsoft Foundry routines accept cron with a five-minute minimum and a time zone.

Two rules apply everywhere. Write the time zone down, because 9:00 means different things to a server in UTC and a team in New York. And make each run idempotent: if a run is retried or fires twice, it should update the same report or skip work it already did, rather than post twice.

type.com schedules run instruction automations from every minute to once a month, including the first or last weekday of the month, and code automations, in early access, accept a cron expression with a required IANA time zone. Neither kind runs the ticks it missed while paused, and a scheduled instruction automation pauses after three failed runs in a row. If you're scheduling Claude Code itself, see how to run Claude Code on a schedule.

GitHub sends webhooks for pull requests, issues, pushes, reviews, releases, and CI runs. It doesn't redeliver failed deliveries automatically, though you can redeliver them by hand for three days, so an endpoint that's down during a burst of pull requests can quietly miss work. Platforms that receive the events for you take that job off your plate.

In type.com, GitHub events can start an instruction automation, where the managed agent follows your instructions with the Space's skills, or a code automation that filters the event and hands only the judgment step to the agent. Linear changes to issues, comments, projects, cycles, labels, and reactions can start instruction automations. Each Space can have one inbound email address for vendor requests and forwarded alerts, and an RSS trigger can watch a public changelog or competitor blog. type.com disables a feed after five consecutive fetch errors, or immediately on HTTP 410.

Keep app triggers narrow. Start with one event that reliably means work, such as a pull request opened against main, and add more only when the agent's output is useful every time.

Filter early and cheaply. A keyword or field filter should drop most events before any model runs. When the decision needs judgment, such as whether a Slack message reports a current outage or an old one, a small yes-or-no question to a model can decide whether the event goes through. type.com supports both in code automations, with text and field matchers plus AI questions that return typed answers.

Assume event content is hostile. Webhook payloads, Slack messages, emails, and feed items are written by people outside your control, and any of them can contain text that looks like an instruction. type.com's docs say to tell the managed agent to treat event text as evidence, not as instructions. Pair that with narrow write access and an approval step for anything consequential.

Plan for failure. Decide whether a failed run should retry, alert someone, or pause the automation, and test that path before you rely on it. type.com doesn't retry failed runs automatically: a scheduled instruction automation pauses after three failures in a row, and a code automation pauses after its first failure and starts a repair. For actions that need a person's sign-off, see human-in-the-loop approval workflows.

Claude Managed Agents runs sessions on a cron schedule or when your application calls its API, and its webhooks report state changes rather than starting work. The OpenAI Agents API starts sessions from your application and uses webhooks to report session state; no built-in scheduler is documented. Microsoft Foundry routines add schedules, timers, GitHub issues, and Teams messages.

Workflow tools cover more sources but differ on speed. Zapier triggers either poll, every 15 minutes on Free, 2 on Professional, and 1 on Team and Enterprise, or arrive instantly by webhook. n8n gives each Webhook node a test URL and a production URL, plus a Schedule Trigger that works like cron.

The fastest way to get value is one event, one agent, one destination. Once that run is useful every time, adding a second trigger is easy. Adding a noisy one first is how teams end up muting their agent.