Skip to main content
A task is one run of an agent. For example, if an invoice agent reviews five invoices, that is five tasks.
Task feed showing the agent's steps and final result for an invoice

A task feed: your request at the top, each step the agent ran (Extract invoice fields, Check the vendor), and the final result.

​
Start a task

You can start a task in a few ways:
  • Start it manually from the agent.
  • Upload a file and ask the agent to process it.
  • Let a trigger start it, such as a new email or a schedule.
  • Start it from an app or API if your team has set that up.
  • Let another system start it with a webhook, such as an EHR that sends an event when a fax arrives.

​
Start tasks from a webhook

Add a Webhook trigger to give the agent a URL. Each JSON event posted to it starts one task, with the event as the task input.
  1. Choose how the sender proves it’s them: a signature made with the sender’s signing secret, a shared header that every event carries, or I don’t have it yet. For a signature, pick the sender (for example Healthie, Stripe, GitHub, Svix or Slack); for any other sender choose Custom and describe how it signs.
  2. Copy the URL and paste it into the sender’s webhook settings.
  3. If you chose I don’t have it yet, open the trigger and add the signing secret once the sender shows it. Until then, events are accepted but start no tasks.
  4. Publish the agent. A new webhook starts tasks only once the published version includes it.
Repeated deliveries of the same event start one task.

​
Custom signatures

When the sender isn’t in the list, choose Custom and describe how it signs, as JSON. Choosing Custom right after a listed sender starts you from that sender’s description. For example, a sender that puts hex(HMAC-SHA256(secret, timestamp + "." + body)) in X-Signature and the Unix time in X-Timestamp:
The message must sign the body, or a digest header that is checked against the body, so a signature can’t be reused with a different body. The spec is checked when you save.

​
Follow the task feed

The task feed shows the agent’s work step by step. You may see:
  • Files the agent read.
  • Data the agent extracted.
  • Lookups it performed.
  • Questions it asked.
  • Files or messages it created.
  • Final results.

​
When the agent asks a question

Sometimes the agent needs help. It may ask because a field is missing, a result is unclear, or your rules require human input.
A paused task showing the agent's question and a Waiting for user input prompt

When the agent needs input, the task pauses with a 'Waiting for user input' prompt. Reply in the box to continue.

When this happens:
  1. Read the question.
  2. Check the document or context if needed.
  3. Reply with the missing information.
  4. The agent continues the task.

​
When a task needs review

Some tasks pause for approval. This is useful for payments, customer messages, system updates, or anything your team wants a person to check first.
A task awaiting review showing a payment approval prompt with Approve, Request changes, and Reject actions

A task awaiting review: the agent explains the decision and pauses with Approve, Request changes, and Reject actions.

Review the result, then approve or reject it based on your team’s process.

​
Common task states

​
Improve the next run

After a task finishes, look for repeated corrections. If you keep fixing the same thing, add a rule so the agent can handle it next time. Learn more in Rules.