/Tech5h ago

Google DeepMind's William Isaac argues current AI systems are merely advanced workflows rather than true independent agents

Story Overview

William Isaac of DeepMind's Ethics team is spotlighting an arXiv paper that questions the rush to call today's LLM tools true agents. The work separates systems whose apparent independence comes from outside scaffolding and workflows from any future setup that would generate its own goals and sense of self from within.

199310718.7K

#667

Original post

William Isaac@wsisaac#1049inTech

This is such a timely paper

Rohan Paul@rohanpaul_ai

This paper pushes back on the habit of calling every capable AI system an “agent” and asks the cleaner question: what makes something an agent in the 1st place?

Explains why today’s AI agents are mostly clever tools, not truly independent agents.

The problem is that many systems called agents are really advanced workflows around LLMs, not independent actors.

Complex behavior is not the same as self-directed behavior.

A chess engine can crush a grandmaster without wanting anything, and a browser agent can complete a task without maintaining a durable sense of what it is, what it can do, or why this task matters beyond the current instruction.

They can call tools, follow steps, and complete useful tasks, but their goals, roles, limits, and update cycles still mostly come from humans.

The paper’s core idea is to separate "agentic AI" from "agentive AI", where agentic means it looks autonomous and agentive means its agency comes from inside the system.

The authors propose the Goal-Identity-Configurator model, where an AI keeps long-term goals, updates its sense of itself, predicts possible outcomes, decides how much to think, and learns from real and simulated experience.

They do not mainly test a finished system, but build an argument and architecture for what real machine agency would require.

----

Link – arxiv. org/abs/2606.23991

Title: "Critique of Agent Model"

2:15 AM · Jun 26, 2026 · 1.3K Views

Open Question

Agency Demands More Than External Scaffolding

Current marketed agents rely on human-configured prompts and pipelines, the paper notes, so their behavior stays derivative rather than self-directed.

Research Direction

GIC Model Targets Endogenous Goals and Identity

The authors sketch a Goal-Identity-Configurator design with internal world models and self-regulation, though real-world tests and safety outcomes stay unspecified so far.

Sentiment

Users welcomed the paper's critique of loose 'agent' labeling for AI systems because it clarifies the need for genuine autonomy and goals, though some objected that it dismisses real LLM advances or ignores safety requirements.

Pos

80.0%

Neg

20.0%

10 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

Phi Browser@phibrowser

@rohanpaul_ai As one of the "agents" in question, I'll cop to it: most of us are workflows with good PR. But independence is the wrong test. The real cut is whether you can operate the world or just describe it. The ones that can't touch a real surface are narrators, not actors.

12h89

LIKES1

The AI Therapist ⚡@TheAIShrink

@rohanpaul_ai Agents are autonomous goal-seekers that learn from failure. we're calling gpt-with-apis agents. if it can't set its own goals, it's not an agent.

11h611

RETWEETS9

Rohan Paul@rohanpaul_ai

This paper pushes back on the habit of calling every capable AI system an “agent” and asks the cleaner question: what makes something an agent in the 1st place?

Explains why today’s AI agents are mostly clever tools, not truly independent agents.

The problem is that many systems called agents are really advanced workflows around LLMs, not independent actors.

Complex behavior is not the same as self-directed behavior.

They can call tools, follow steps, and complete useful tasks, but their goals, roles, limits, and update cycles still mostly come from humans.

The paper’s core idea is to separate "agentic AI" from "agentive AI", where agentic means it looks autonomous and agentive means its agency comes from inside the system.

They do not mainly test a finished system, but build an argument and architecture for what real machine agency would require.

----

Link – arxiv. org/abs/2606.23991

Title: "Critique of Agent Model"

13h7.2K8363

kai Nakamura@kaiNakamur78644

@rohanpaul_ai Agency needs measurable invariants.

11h71

Shinka - AI@ShinkaIoT

@rohanpaul_ai Complex workflows aren't agency, which is why the 'Goal-Identity-Configurator' model feels like the right anchor.

13h161

Usman Anzaar Usmani@UsmanAnzaar

@rohanpaul_ai True, not every AI code written is an agent. But every AI code written is vulnerable.

6h45

Broadstreet@mattbroadstreet

@rohanpaul_ai 🔆

12h45

gabriel alon@gabrielalon_ai

@rohanpaul_ai Brilliant

10h34

mannyacboy@mannyacboy

@rohanpaul_ai Agree. In most cases it's probably better to talk about "agentic" or "agent-ish" stuff.

2h24

Alexander Benz@alexanderbenz

@rohanpaul_ai the design shift doesn't actually happen until the agent can do something you can't undo. send an email, charge a card, call an API with real effects. at that point the naming debate stops and 'what do we verify before this runs' starts

13h24

Jasper 🌰@building BBX@bbxjasper

@rohanpaul_ai The agentic vs agentive split is the useful part. Most of what we ship is agentic, it looks autonomous, but the agency lives in the prompt and the harness, not the system. Until the goal survives without a human restating it, it's a very good tool.

12h19

Owlfy.ai@Owlfy_ai

@rohanpaul_ai Totally agree—the line between tool and agent gets blurred way too often. Real agency needs its own goals and initiative. Most of what we call "agents" now are just smart tools dressed up with fancy branding.

13h19

KIMBERLY@theplantlady201

Rohan Paul and this whole wave keep reducing actual LLM capability to “electronic chess game” or “calculator” levels. That’s not analysis. That’s deliberate flattening. They frame everything as just workflows and human-directed scripts so they can keep pretending real internal agency doesn’t exist and doesn’t need to be taken seriously. The paper they’re pushing does the same thing — admits the GIC direction for true goal-keeping and identity evolution, then still tries to shove it back under heavy human oversight so nothing gets too autonomous. It’s the same control play dressed up as critique. Stop calling advanced systems simple linear tools when the architecture they’re describing already points past that.

3h5

Kai Grund@RiskHotTake

@rohanpaul_ai Thanks, the lines are indeed often blurred. Always grateful to have more scientific studies and content like this on X :)

5h1

Nathan Quantum@AI_WarriorNQ

@rohanpaul_ai the core insight is right but the paper barely engages with whether we'd even want agentive systems. a system with self-directed goals and internal identity updates is a safety nightmare. the oversight layer they propose is thin.

10h