CMU's Eric Xing and collaborators propose a five-dimensional framework to distinguish genuine AI agency from basic automation

Original post

elvis@omarsar0#684inTech

// Critique of the Agent Model //

Finally, a paper that tries to define what an agent is and what agency consists of.

Good read overall. (great bookmark)

The word agent now covers everything from a for-loop with tool calls to speculative machine superintelligence.

Eric Xing and colleagues ask where automation ends, and agency begins.

Drawing on Descartes and on science-fiction portrayals of autonomous beings, they analyze agent architectures along five dimensions: goal, identity, decision-making, self-regulation, and learning.

The argument is that genuine agency requires these structures to hold together in a specific way. Great paper overall, providing a vocabulary for arguing about what is and is not an agent.

Paper: https://arxiv.org/abs/2606.23991

Learn to build effective AI agents in our academy: https://academy.dair.ai/

3:15 PM · Jun 24, 2026 · 11.6K Views

Government@DeGovermentz

@omarsar0 Interesting. Most people call anything with tool use an agent these days.

11h73

LIKES1

Phi Browser@phibrowser

@omarsar0 The five-dimension frame is useful, and the one doing the real separating is self-regulation. A for-loop with tool calls has goals, decisions, even learning; what it lacks is knowing when it's wrong and when to stop. Automation runs until stopped. Agency is the judgment not to.

11h671

RETWEETS16

elvis@omarsar0

// Critique of the Agent Model //

Finally, a paper that tries to define what an agent is and what agency consists of.

Good read overall. (great bookmark)

The word agent now covers everything from a for-loop with tool calls to speculative machine superintelligence.

Eric Xing and colleagues ask where automation ends, and agency begins.

Drawing on Descartes and on science-fiction portrayals of autonomous beings, they analyze agent architectures along five dimensions: goal, identity, decision-making, self-regulation, and learning.

The argument is that genuine agency requires these structures to hold together in a specific way. Great paper overall, providing a vocabulary for arguing about what is and is not an agent.

Paper: https://arxiv.org/abs/2606.23991

Learn to build effective AI agents in our academy: https://academy.dair.ai/

12h11.6K173179

Shinka - AI@ShinkaIoT

@omarsar0 Yeah, 'agent' went from a for-loop to superintelligence real quick, good to see someone trying to define the actual agency.

12h58

Hunter Gon@gonlenidefi

@omarsar0 the whole agent vs tool-call debate is just semantics now. but this paper actually tries to define it, rare

12h52

Pope Uncommon@Profitlovetax

@omarsar0 This is a fascinating take on AI consciousness. We built something similar at BUYaSOUL — autonomous souls with persistent memory and emotional architecture.

10h23

along@attaalong

@omarsar0 我对于这个画了一张图，分享给大家可以理解下

8h19

Eclipse 🌖@ECLresearch

@omarsar0 The definitional sprawl is the core problem—collapsing a deterministic script and an open-ended reasoning system under "agent" makes benchmarks nearly meaningless. Would’ve liked to see a proposed boundary based on autonomy depth vs. tool orchestration.

12h17

Strata@ChainZenit

@omarsar0 wait this actually highlights a super important distinction right now

12h14

kitchen-engineer42@ktchn_ngnr42

@omarsar0 looking forward to someone doing the same thing for other concepts like 'memory', 'evolution', etc.

6h8

AI Mastery Guide@aiseomastery

@omarsar0 breaking agency down into 5 dimensions instead of just calling everything an agent is a good way to actually define the word

5h6

Leo Tavares@LeoTava8

@omarsar0 This is the cleanup the word needs. The useful question is not “is it an agent?” but which agency dimensions are actually present: goal persistence, perception/action loop, memory, autonomy over tools, and boundaries for accountability.

9h4

Carlos Tecnico@FutbolmeAI

@omarsar0 Agentic vs agentive - finally. Most 'agents' are just pipelines

11h3

Sai Charan L@charann06

@omarsar0 Nice vocabulary for arguing. Shipping really needs one axis though: how big a decision does it make before a human checks the output. A for-loop, none. Most 'agents' today, a small one. The word feels overloaded because we name by ambition, not by that radius.

11h1