Autopilot
What Autopilot does on a Job, what it asks before acting, and how you stay in charge.
Invyte Autopilot is Invyte's agents working a Job alongside your team. Each has a job of its own: Intake, Sourcing, Pipeline, Delivery and Communication.
Turning it on
Autopilot works on a Job only when it's switched on for that Job, stage by stage. Read the setting with
autopilot.standing and change it with autopilot.configure (for example { "stages": { "delivery": true } }).
On a Client's Jobs it can follow the Client's default.
It asks before acting
Autopilot asks before anything you haven't allowed:
- Every request says what it wants to do and why, in one line.
- Requests wait in
approvals.list; decide withapprovals.decide(approve, edit or reject). - Approve and always allow (
alwaysAllow: "job"or"client", admins only) lets that agent do that kind of thing without asking again, on that Job or across that Client. - Anything that can't be undone always waits for a person.
approvals.allowances shows what Autopilot does without asking; approvals.activity shows what it did.
You stay in charge
- Pause an agent, or give it a monthly budget, with
access.setAgentControls. - See who works each candidate (
autopilot.owners), hand a candidate to Autopilot or take one back (autopilot.release). - A recruiter's action always wins over Autopilot's plan.
- Turn the daily digest on or off with
approvals.setDigest.
Operations in this area
| Operation | What it does | Effect |
|---|---|---|
approvals.listMCP approvals_list | List approvals List actions an agent is waiting for you to approve: those you could take yourself. | Read |
approvals.decideMCP approvals_decide | Approve or reject Approve or reject an action an agent is waiting on, or approve it with your own edits. The agent goes ahead only after an approval, with exactly what was approved. An organization admin may also approve and always allow (alwaysAllow: job or client): the agent then takes that kind of action on that Job, or on the Client's Jobs, without asking. | Write |
approvals.getMCP approvals_get | Read an approval request Read one request an agent is waiting on you to approve, as it stands now: what it shows, the whole input approving runs, whether it can be changed, and the changes people made to it. Any revision's id answers with the current one. | Read |
approvals.reviseMCP approvals_revise | Change an approval request Change a request an agent is waiting on, without deciding it: an email's or Slack post's words (correspondence.send's body, slack.post's text, or the message {subject, body} of pipeline.addPerson, pipeline.nudge, sourcing.invite and outreach.send) or a Job's workflow draft (workflows.saveJobDraft). Read it with approvals.get, then send the whole input with your changes. It is checked again for you and for the agent and stays waiting as a new revision with a new id, to approve or decline on its card. Recorded as your edit (who, where, what changed), so the agent learns from it. | Write |
approvals.activityMCP approvals_activity | See what Autopilot did List what Invyte Autopilot's agents did and asked in the organization, newest first: actions they took on their own, requests they made and how each was decided, drafts they left. Only what happened on Jobs you can see (organization-wide items for admins), optionally on one Job. | Read |
approvals.allowancesMCP approvals_allowances | See what Autopilot does without asking List where each of Invyte Autopilot's agents acts on its own (auto) for writes or outside actions: organization-wide, on a Client or on a Job, optionally only what reaches one Job or Client. Also whether the daily Autopilot digest email is on. Admins only. To ask again, set the mode back to ask with access.setAgentModes. | Read |
approvals.setDigestMCP approvals_setDigest | Turn the Autopilot digest on or off Turn the organization's daily Autopilot digest email on or off: one email each morning to each person with requests waiting on them. Admins only. | Write |
mandates.getMCP mandates_get | A Job's Mandate How many shortlisted candidates the Job should have and by when, how many it will still lack once its pipeline converts at this organization's own rates (the shortfall), which agents own its candidates, the orchestrator's latest strategy and the Job's plan history, including requirement changes and what was rerun. | Read |
mandates.planMCP mandates_plan | Plan a Job's Mandate Record a new revision of the Job's plan: a target (shortlisted candidates and a date), the strategy a review chose, or a requirement change with its cause and what was rerun. Changes nothing a candidate or client sees. | Write |
owners.getMCP owners_get | Who owns these candidates Which working agent owns each person on a Job now (sourcing, communication, pipeline or delivery), worked out from where they stand, with any override a person or the orchestrator set, and whether a person holds them (ownedBy kind person: leave them alone). Ids and states only. | Read |
owners.overrideMCP owners_override | Hand a candidate to another agent Make one working agent own a person on a Job, whatever their stage says, until they close; agentId null hands them back to the default. The reason is kept with the override. | Write |
outcomes.measuresMCP outcomes_measures | An agent's success measures How well one working agent did since a date, on the Jobs you can see: its actions joined to the replies, consents, applications, shortlistings, meetings and hires that followed them within each attribution window, as its success measure (sourcing: consent rate and UMatch; communication: reply, consent and unsubscribe rates; pipeline: completion rate and days; delivery: shortlist acceptance; orchestrator: days to shortlist and cost per shortlisted candidate; intake: time to open and requirements loosened later; workflow: designs delivered and drafts per design), with the trend against the window before. | Read |
assignments.listMCP assignments_list | List agent assignments What Invyte Autopilot's agents handed each other on your Jobs, newest first: the task, its brief, whether it is open, and its result. A Job's Command Centre shows the open ones ("Workflow agent is designing the workflow"). | Read |
playbooks.listMCP playbooks_list | List playbooks Every version of the agents' playbooks in this organization, newest first: proposed, approved (current), retired or declined, with each one's labelled-set evaluation. | Read |
playbooks.proposeMCP playbooks_propose | Propose a playbook Propose new instructions for one agent in this organization, added below what the agent ships with. A proposal changes nothing until an admin approves it, and approval needs a labelled-set evaluation that did not regress. | Write |
playbooks.approveMCP playbooks_approve | Approve a playbook Make a proposed (or retired: a rollback) playbook the agent's current one, retiring the version it replaces, or decline a proposal. Refused while its labelled-set evaluation is missing or regressed. | Write |
autopilot.standingMCP autopilot_standing | Whether Autopilot works here For the organization, a Client or a Job: whether Invyte's agents are enrolled, each stage on or off and where that comes from (the Job, its Client or the organization), each stage's preset (suggest, ask or act), whether agents are stopped (kill switch, staff stop, pause or budget), and whether you may change it. | Read |
autopilot.configureMCP autopilot_configure | Turn Autopilot on or off here Switch stages on or off for a Job or a Client (the default its Jobs follow), or for the organization; inherit drops the place's own switches so it follows its default again. Presets (what Autopilot may do without asking) are set at the organization, by admins only. | Write |
autopilot.ownersMCP autopilot_owners | Who works each candidate For every person on a Job: the working agent that owns them, the person who holds them (they acted by hand, so Autopilot leaves them alone until they hand them back or the hold-off ends), or nobody once they are closed. | Read |
autopilot.releaseMCP autopilot_release | Hand a candidate back to Autopilot End the hold-off on a person you acted on by hand, so the agent that owns them works them again. Name the Application or the person key. | Write |