Tasks and runs
A task is a unit of work given to an employee. A run is one execution of that task. Each run is made of steps, and Orqane saves checkpoints along the way so a run can pause and resume, even over several days.
| Field | What it sets | Example |
|---|---|---|
| Goal | What the employee must achieve, in plain language. | Qualify the 40 new leads from yesterday. |
| Inputs | Files, records, links or data the task starts from. | A CSV export, a HubSpot list |
| Assignee | The employee, or the team, that does the work. | Ada |
| Owner | The person or employee who asked, and receives the result. | Camille Martin |
| Deadline | When the result is due. The employee plans around it and warns the owner if it will be late. | Tomorrow, 09:00 |
| Priority | Low, Normal, High or Urgent. Higher priority tasks start first. | High |
| Delivery | Where and how the result is delivered. | Post a table in #sales |
| Trigger | What created the task. Set by Orqane. See Triggers. | Slack mention |
A task has one status: Open, In progress, Blocked, Done or Canceled. It’s Blocked when its run waits for an approval or for input from a human.
The engine loop
Section titled “The engine loop”When a run starts, the employee works in a loop until the goal is met, the deadline passes or a limit is reached.
Plan
The employee reads the goal, retrieves relevant memory, picks the skills to apply and writes a plan. You see the plan at the top of the run in Activity.
Act
It performs the next step: a tool call, a message, a payment or a hand-off. Before each action, Orqane checks the autonomy level, policies, guardrails and budget.
Observe
It reads the result of the step: the tool’s output, an error, a reply.
Reflect
It checks progress against the goal and the plan. It updates the plan if needed, then goes back to Act, or delivers the result.
Each pass through the loop is recorded as steps you can replay. See Activity and traces.
Sub-tasks and sub-agents
Section titled “Sub-tasks and sub-agents”Large tasks are split:
- Sub-tasks go to other employees, through a hand-off or a manager employee. Each sub-task has its own run, under the receiving employee’s autonomy level and budget.
- Sub-agents are short-lived helpers the employee starts inside its own run, for example to research 10 accounts at once. A sub-agent has the same skills, tools and limits as its parent, and it can’t create sub-agents of its own. Its cost counts toward the parent’s budget.
Parallel steps
Section titled “Parallel steps”When steps don’t depend on each other, the employee runs them in parallel, such as reading five web pages or querying three systems. Up to 10 steps run at once per run by default. Steps that need approval are never run in parallel with the action they unlock.
Long-running runs
Section titled “Long-running runs”A run can last minutes or days. A run that chases late invoices, for example, may send reminders and wait for replies over two weeks. While it waits, it uses no credits. It wakes up on an event (a reply, an approval, a webhook) or at a scheduled time.
Checkpoints and resume
Section titled “Checkpoints and resume”Orqane saves a checkpoint after every step that changes something, and before every approval. A checkpoint holds the plan, working memory and the results of all steps so far.
A run resumes from its last checkpoint when:
- an approval is answered,
- a human replies to a question,
- a paused employee is resumed,
- a failed step is retried after an outage.
Completed steps are never repeated. A run that already sent an email doesn’t send it again on resume.
Run statuses
Section titled “Run statuses”| Status | Meaning |
|---|---|
| Queued | Waiting to start, for example because the employee is at its concurrency limit. |
| Running | The engine loop is active. |
| Waiting for approval | Paused at a checkpoint until an approval is answered. |
| Waiting for input | Paused until a human or an external system replies, or a scheduled time is reached. |
| Paused | Stopped by a member or because the employee was paused. |
| Completed | The goal was met and the result delivered. |
| Failed | The run stopped on an error it couldn’t recover from. |
| Canceled | A member canceled the run, or the employee was retired. |
Retries and timeouts
Section titled “Retries and timeouts”- Retries. A failed tool call is retried up to 3 times, with growing delays. Steps classed Communicate, Spend or Irreversible are only retried when the tool confirms the action didn’t happen.
- Step timeout. A tool call that doesn’t answer in 5 minutes fails, and the employee decides whether to try another way.
- Run limits. Each run has a maximum of 200 steps and a maximum duration of 30 days by default. Change them on the employee’s Settings tab. See Limits.
When a run fails, its owner and the employee’s human manager are notified. Click Retry from checkpoint to resume, or Retry from start to start a new run.
Take over a run
Section titled “Take over a run”You can take control of any run of an employee you manage. Open the run in Activity and click Take over. The run pauses and you can:
- edit the plan or the next step,
- perform a step yourself, such as writing the email by hand,
- give the run back with Hand back, and the employee continues from there,
- finish the task yourself with Mark as done.
Your actions are recorded in the run, and the employee learns from them.