Model plus harness
Imagine an assignment: new data has arrived in the source spreadsheet, and the report, charts and slides that depend on it need updating. The work is not done when the agent has produced text or called a tool. It is done when the linked documents match the assignment — and some commitments may still be waiting for someone's approval.
The authors propose looking at a work agent as a system: the model understands the task and proposes actions, while the harness — the execution environment — manages context, tools, memory, execution and recovery from failures. Together they operate in a stateful workspace, and the result is checked against observable signs.
Two axes of conditions
Tasks differ along two independent axes. Duration of responsibility: does the task close at delivery, or does it stay open until the next stages and events. Openness: is everything known in advance, or do objects, links and rules have to be discovered along the way.
Diagram based on the authors' framework. The axes are not a difficulty scale but different questions about the task.
Five forms of work
- Conversation. The result is information.
- Atomic action. A single local change.
- Multi-step task. Dependent actions toward one goal.
- Process. Stages and handoffs between participants.
- Workspace. You have to find and link the right objects yourself.
A cycle of five capabilities
Perception finds the right objects and links, reasoning weighs goals and constraints, planning builds dependencies, execution acts through tools, and verification checks the result against the assignment and triggers a fix or replanning. These are functions, not necessarily separate modules: they can be performed by the model, the harness or both together.
How to evaluate
An agent can be improved in three ways: fine-tune the model, improve the harness — interfaces, processes, execution control, memory and skills — or develop both together. And the authors propose evaluating along three separate dimensions: whether the work is accepted against the agreed requirements, whether it brings value in real use, and how much human effort and compute it took.
What it means in practice
When deploying a corporate agent, the acceptance criteria and the way the result is checked have to be set before development, not after. It is worth improving not only the model but also the harness: it is often what decides whether the work gets finished. A caveat: this is a survey and a framework, not a new agent with experiments — the authors leave measuring its effect to future research.