Assignments

An assignment is a job described once and run many times: a name, the instructions the agent follows each run, and the tools it needs. Assignments belong to an agent and live on the agent’s page.

Assignments are the step from delegation to routine. A conversation gets you one answer; an assignment gets you the same job done every Monday morning, or every time another system calls a webhook, without anyone remembering to ask. Each run is a session — the full record of what the agent did — so a recurring job builds its own audit trail as it goes. Most of the standing work in a workspace ends up here: the weekly report, the maintained tracker, the triage of whatever arrived overnight.

Creating an assignment

On the agent’s page, open Assignments and choose Add Assignment. An assignment has four sections:

Instructions — the job prompt: what the agent should accomplish each run. The AI-assist button drafts instructions from a short description, using the agent’s personality as context. Be explicit about the deliverable; “produce a document with a summary table and week-over-week changes” beats “summarize the pipeline.”
Schedule — the triggers: every assignment can run on demand, and can also run on a schedule or from a webhook. See Schedules, webhooks, and triggers.
Options — parameters the run receives, skills and integrations the agent should always use (by default the agent decides), an optional deadline, and Ultrawork mode for maximum-effort planning on the hardest jobs.
Hooks — a read-only view of any follow-up actions attached to the assignment, which run automatically after it completes successfully.

Save the assignment, then use the Run now button on its row to start a run immediately (the button enables once the assignment is saved).

A complete example

Here is the assignment the weekly report guide builds, written out section by section.

Name — Weekly Pipeline Report.

Instructions —

Every Monday, pull the open opportunities from the CRM. Produce a document with a summary table of pipeline by stage, the top five at-risk deals with a one-line reason each, and week-over-week changes in total value. Call out any deal that changed stage since last week. Keep the whole thing to one page — it gets read on phones before the Monday standup.

Schedule — Weekly, Mondays 6:00 AM, America/New_York.

Options — the CRM integration is listed under integrations the agent should always use, so no run skips it; no parameters, no deadline, and Ultrawork mode stays off for a routine job like this.

Save, press Run now, and watch the first session. If the report isn’t right, refine the instructions and run again; the schedule takes over once you’re happy with the output.

From conversation to assignment

If you find yourself asking an agent for the same thing repeatedly, turn the conversation into an assignment: the Make recurring button on a session distills it into a draft — name and instructions pre-filled from what the agent actually did — which you then review and schedule. This is often the easiest way to write good instructions: a real conversation shakes out the ambiguities before anything lands on a schedule.

The assignment list

The Assignments card groups assignments by cadence — Hourly, Daily, Weekly, Monthly, Custom, One-Time, On-Demand — with a Completed group for finished work. Each run appears in the agent’s sessions, so the full history of any assignment is always an audit trail away: what ran last Monday, and how it differed from the Monday before.

Assignments can also surface where the team tracks its work: a board can adopt an existing assignment as a card and manage its schedule and status from then on.

Pausing

Each assignment has a pause switch that stops its automated triggers, and the agent has a master “Pause all automated runs” switch that quiets everything at once. Manual Run now keeps working either way — pausing stops the automation, not the job. Reach for the per-assignment switch while you rework one job’s instructions; reach for the master switch when the whole agent should go quiet, over a holiday week or while you reorganize its workload.

When something goes wrong

A run that hits a blocker — expired credentials, missing data, an ambiguous instruction — escalates to your Inbox with a priority rather than failing silently. Fix the blocker, reply, or rerun. The practical consequence is that you don’t need to check whether the Monday report ran: if it did, the deliverable is there, and if it couldn’t, the Inbox says so and tells you why.

Patterns that work

Standing report — a weekly or monthly schedule with a document deliverable: pipeline, support volume, campaign performance. The weekly report guide is the worked version.
Maintained surface — the job is to keep a data table current rather than to produce a fresh artifact each run; see the agent-maintained table guide.
Event-driven work — a webhook trigger, so a run starts when a ticket, form submission, or deploy happens in another system; see triggers.
On-demand playbook — no standing trigger at all: a codified job you start with Run now whenever the situation calls for it.