Concept: Agent-Shaped Work
The framework for identifying, estimating, and structuring business tasks to determine their suitability for AI agent execution. Understanding what work looks like when it is “agent-shaped” prevents the underutilization of AI capabilities (such as the openclaw idle agent phenomenon) and avoids over-engineering simple tasks.
The Agent Test (Four Primitives)
To evaluate any task on your desk in about a minute, estimate the following four dimensions:
- Size: Is the task bigger than what one agent can hold at full quality? (e.g., a calendar or a single document fits in a context window; a pile of 1,000 documents exceeds it, requiring multi-agent scaling).
- Independence: Can the parts of the task be done in parallel without knowing what the other parts did? (e.g., reading separate documents splits well; complex monolithic coding may not).
- Separation of Concerns: Do any parts of the task need to be done by different minds? This prevents parts from “poisoning” each other (e.g., an auditor cannot keep the books; peer review requires a separate reviewer; fresh eyes are needed for drafting vs. editing).
- Checkability: Is checking an answer much cheaper than producing one? (e.g., a test suite, an exit code, or a source document reference). If checkability is cheap, multi-agent scaling is highly effective. If checking is expensive, multi-agent value tops out quickly (around 100 attempts).
Academic & Practical Foundations
- Stanford 2024 Study (The Token Law): Giving a cheap model 250 attempts at a bug instead of 1 improved success from 15.9% to 56%. However, scaling to 10,000 attempts (where a correct answer exists 95% of the time) requires an automatic checker (evals) to find the correct answer. Without a checker, majority voting and reward models stall out at 100 attempts.
- Anthropic Study: Token spend (how much the system is allowed to think) explains 80% of the difference between a good and bad run. A team of agents is how you spend more tokens than a single agent can hold to tackle hard problems.