Ask an agent to sort out a delayed order and it starts at once. It looks up the order, drafts a reply, perhaps books a courier. Whether the customer has the parcel this week is another matter, because that depends on things the agent was never asked to hold together: the delivery and refund rules, the courier’s latest update, who has to approve the extra cost, and what happens when the customer changes the address halfway through.
That is the gap between doing tasks and reaching goals. Most AI today is good at the first. A goal asks for more: a plan, a view of the options, a decision someone is accountable for, work done in real systems, a check that it worked, and the patience to carry on when something changes.
Duplex is built for the second. It is the Agent Operating System powered by Runtime Intelligence: you set the goal, and Duplex coordinates everything the goal needs, within your rules, until the result is verified.
A goal arrives, and Duplex brings in the context and sets out a plan.
A goal is more than a list of tasks.
A task has an end you can see from the start: send the email, update the record. A goal has an end you have to reach: the order delivered, the supplier set up, checkout working again. Between the two sits everything that makes work hard. The facts are spread across systems. There is more than one way to get there, and each costs something different. Some steps need a person’s approval. And things change while the work is under way.
When agents handle tasks one at a time, all of that falls to people. Someone gathers the context, weighs the options, chases the approval, checks the outcome and starts again when a fact changes. The agent saves a few minutes on each task, and the goal still takes as long as it always did.
Coordination is where the time goes, so coordination is what Duplex takes on. It brings the parts of the work together around the goal, so the plan, the decision, the work and the check become one piece of work instead of a relay between people and tools.
Everything the goal needs, working as one.
Before anything runs, Duplex brings in the context that matters: the history of the case, the facts it depends on and the rules that apply. It sets out a plan and plays each option forward, so the likely outcome of each is clear before time or money is spent. Where a case is new, it works out an approach from what it already knows, and checks it before anyone relies on it.
While the work runs, each step is taken for a reason. Duplex goes ahead with what your rules allow and hands the rest to a person, with the options and their outcomes laid out. Every model call and tool action is checked against your rules before it runs, and the work itself happens in your systems, through the tools and agents your teams already use.
Over time, the work keeps its shape. A long job stays one job, with its goal, its plan and its progress. A pause, a hand-off or a changed fact costs a moment rather than the work. And once the result has been checked against the goal, what worked becomes a skill for the next goal like it.
None of these parts is new on its own. What is new is that they work as one system around your goal, instead of as separate tools that someone has to wire together and watch over.
Your rules, your approvals, one record.
Coordination is only worth having if you can trust it, and trust starts with control. You decide the goals, the rules and the actions that need a person. Duplex holds every agent to them in every session, and the approvals it asks for go to the people you choose.
Every step leaves a record: what was decided, on which facts, under which rule and with whose approval. When someone asks why the express courier was booked, the answer is there, with the manager’s approval beside it.
The same record is what lets work move between people. A colleague who picks up the case the next morning, in another tool, starts from the same goal, plan and progress, not from a summary pasted into a chat.
Learning is one part, not the whole story.
Duplex learns from the results your team confirms, and that matters: the verified way to resolve a delayed order is ready for the next one, and it stays yours when you change models. But learning is one part of reaching goals, not a substitute for the rest. An agent that learns without planning, checking and keeping to your rules only repeats its mistakes faster.
So Duplex puts learning where it belongs, after the result has been checked. What your agents reuse is what reached the goal, within your rules and with the approvals it needed.
Give Duplex the goal, not the steps. It brings together the plan, the options, the decision, the work, the check and the patience that goals take, and your agents finish what they start.