The Three Dimensions of an Agent
Score any 'agent' on autonomy, complexity, and natural interaction — each a spectrum
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 93%
A practical definition and design lens for agents that strips away the hype. Three product principles distinguish a real, good agent from a one-shot AI feature: increasing autonomy (delegating higher-order tasks along a spectrum, not a binary), complexity (multi-step goals, not single-shot outputs), and natural interaction (talk it through, jump on a meeting with it), plus the property that it works asynchronously while you aren't.
Origin
Aparna Chennapragada's computer-scientist framing of agents, developed while leading agent work at Microsoft (e.g. a researcher agent 'made for work').
Core principles
- 01Agents are tools — stochastic models under the hood — not magic
- 02Autonomy is a spectrum of delegation, not a zero/one switch
- 03Good agents deliver new insight, not just time saved
- 04Asynchrony ('works when you're not working') is a defining property
How to run it
- 1
Rate autonomy / delegation
Ask how high-order a task you can hand off. As deep-reasoning unlocks arrive you can delegate more — from fine-motor handholding to 'here's my goal, go make it happen.'
Pro tip Treat autonomy as a dial to raise gradually as the model's reasoning improves, not a fixed setting.
- 2
Rate complexity
Distinguish one-shot tasks (summarize this, generate this image) from complex, multi-step ones (build me this prototype, help me nail this meeting). Agents earn the name on the complex end.
- 3
Rate natural interaction
Beyond chat: can you jump on a meeting with the agent, talk it through, and point it at what you want done differently? Richer interaction modes mark better agents.
- 4
Require asynchrony
A true agent runs while you're not sitting in front of it — it works when you're not working.
Pro tip Design for hand-off-and-walk-away, not for constant supervision.
In the wild
Chennapragada asked her agent to review who would be in an upcoming leadership meeting, surface their views on the topic, and suggest a persuasion approach.
→ It didn't just save time summarizing — it fired synapses she didn't have and delivered genuinely new insight, illustrating the payoff beyond one-shot AI.
Common mistakes
Treating autonomy as binary
Framing an agent as fully autonomous or not, instead of choosing a delegation level appropriate to the task and the model's current reasoning.
Calling a one-shot feature an agent
Labeling single-shot summarize/generate features as agents when they lack complexity, autonomy, and asynchrony.
Is it for you?
Best for
Product builders defining, scoping, or evaluating AI agent products
Not ideal for
Simple one-shot AI features where agent framing over-complicates the design
From the transcript
“when I think about agents, I think about three things. one is is an increasing um level of autonomy”
“it's not just a oneshot, hey create this image or do this thing or summarize a document. It's you know build me this prototype”
“the third thing I would say is asynchronous. It works when you're not working”
“the autonomy, the complexity, and the natural interaction are at least product principles that will shape really good ones, good agents”
From the episode
Microsoft CPO: If you aren’t prototyping with AI, you’re doing it wrong
Aparna Chennapragada