Work Agents: Systems, Capabilities, and Optimization
The framework provides a way to think about designing agents that aim to reliably complete multi‑tool digital work.
In short. The authors propose a system‑level perspective that treats work agents as foundation models plus runtime harnesses operating in stateful workspaces. They define dimensions of work persistence and openness, classify work into five forms, and outline five capability families and optimization strategies.
What it changes
Existing research lacks a unified perspective on how agents reliably accomplish delegated tasks, making it difficult to ensure end‑to‑end reliability and to compare different approaches.
The paper frames agents as systems that combine a model with execution harnesses to act in persistent, possibly open, workspaces. It introduces two axes—persistence (how long work state is kept) and openness (access to external tools)—to categorize tasks. From this, it extracts five capability families needed for dependable execution and suggests joint model‑and‑harness optimization.
Think of it like this: It’s like defining a kitchen’s layout, appliances, and recipes to ensure a chef can reliably prepare a multi‑course meal.
How it works
- Define work dimensions
Identify work persistence (state retention) and work openness (tool access) for the target task.
- Classify work forms
Map tasks into five forms based on scope and organization using the two dimensions.
- Identify capability families
Extract five core capabilities (e.g., planning, tool use, error recovery) required for reliable execution.
- Optimize model and harness
Explore separate and joint optimization of the foundation model and its runtime harness to improve each capability.
- Evaluate outcomes
Apply work‑centric metrics that measure verifiable outcomes in stateful workspaces.
Keep in mind
- The framework is conceptual and has not been empirically validated on real agents.
- It assumes clear definitions of persistence and openness, which may blur in practice.
- Future work must test joint optimization strategies in open, dynamic environments.
What it means for you
Background only: the paper offers a strategic, system‑level view without presenting concrete ROI metrics.
Words worth knowing
- foundation model
- Large pretrained AI model used as a base for many tasks.
- work agent
- System that combines a model and execution code to perform digital work.
- work persistence
- How long the agent retains state about a task.
- work openness
- Degree to which the agent can access external tools or environments.
- stateful workspace
- An environment that keeps track of intermediate results and context during a task.
Written by Deltaday from the paper’s text and checked against it. A simplified explanation: for exact methods and caveats, read the original.
Earlier papers of the week
BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
New benchmark shows top AI models fail over a third of the time on real business tasks due to process errors.