The work that matters most rarely fits in one sitting. Onboarding a customer takes a week. Closing the books takes days. A migration, an audit or a supplier review runs for weeks, with people, systems and facts changing all the way through.
This is where AI agents struggle most. A session ends and the thread of the work goes with it. A fact changes and the plan carries on as if it hadn’t. A step goes wrong and nobody notices until the end. Long work needs something agents rarely have: a hold on the goal that outlasts every session.
Long Horizon is how Duplex gives them that hold. Long work becomes one job, with its goal, its plan, its rules and its progress, carried through pauses, hand-offs and change until the goal is reached.
The goal is set, and Duplex lays out a plan your team approves.
One job, from the first step to the goal.
Set up in Duplex, long work stops being a string of conversations. It becomes one job, with a goal, a plan your team can read and approve, the rules it runs under and a record of every step taken and every check made. Anyone can ask where it stands and see how far it has come and what comes next.
That changes who has to remember. Without it, the people around the agent hold the thread: what was decided on Monday, which approvals came through, what is still open. With Long Horizon, the job holds the thread, so people can look after the decisions rather than the bookkeeping.
It also changes what counts as progress. A step counts towards the goal only once its result has been checked. Long work cannot look finished before it is, and the progress you see is progress that has been verified.
Plans that stay true while the world changes.
Long plans go stale. The supplier sends a new certificate, a price moves, a system is taken down for maintenance. A plan written on Monday can be wrong by Wednesday without anyone noticing, and an agent that follows it faithfully does the wrong thing well.
Before each step of a long plan runs, Duplex checks that what the step depends on still holds. When it doesn’t, Duplex revises the rest of the plan from the work already verified, rather than starting over or pressing on regardless. When a step has to go beyond the plan or your rules, the people you choose approve the exception, and the approval is on record.
Duplex also watches the shape of the work. It notices when a job stops moving towards its goal, goes round in circles or runs past the budget you set, so a stuck job is spotted while there is still time to change course.
Interruptions cost a moment, not the work.
Real work is full of interruptions. Sessions end, people go home, a colleague picks up the case in another tool, a model is swapped for a newer one. Each of these used to mean explaining the work again from the beginning, or finding that a half-finished step had quietly been dropped.
With Long Horizon, Duplex takes the work up where it stopped, with its goal, its plan, its decisions and its approvals. Safe points marked before risky changes mean long work never has to start again from the beginning, and the job can be picked up by another session, agent, app or model.
None of this depends on one person staying at their desk. The job carries its own context, so whoever picks it up starts informed.
What long work leaves behind.
When a long job closes, Duplex settles what it learned. The steps that led to the result are credited. What was learned along approaches that were dropped is withdrawn, so a dead end on Tuesday doesn’t shape next month’s work.
What did work becomes part of what your agents can reuse: the verified way to onboard a supplier, close a period or move a system. Long Horizon carries the work to the goal, and Skill Fabric carries what worked into the next job.
Long work is where AI agents could do the most, and where they have been trusted least. Long Horizon makes it work you can hand over: one job, with its goal held, its plan kept true, its progress verified and its interruptions absorbed, until the goal is reached.