TL;DR: On June 22, 2026, Eric Xing, Mingkai Deng, and Jinyu Hou published Critique of Agent Model on arXiv (2606.23991). The paper draws a sharp line between agentic systems — where competence lives in external scaffolding (workflows, tool APIs, permission hooks) — and agentive systems — where goal, identity, decision-making, self-regulation, and learning are internalized. It proposes GIC (Goal-Identity-Configurator) as a general-purpose architecture combining hierarchical goals, evolving identity, world-model simulation, learned self-regulation, and self-directed learning, with explicit attention to auditability and human oversight.
Why this paper matters now
Every vendor ships an "agent" in 2026. Claude Code, Cursor, Codex, Copilot — the label is everywhere. At the same time, headlines ask whether AI will develop machine agency against humans.
Critique of Agent Model cuts through both hype and panic with a simpler question: What is an agent? What constitutes agency?
The authors argue we need a precise boundary between automation (prescribed tasks, engineered loops) and agency (structures that let a system operate in the open world with genuine autonomy). That boundary matters for building capable systems and for evaluating what is worth fearing.
| Detail | Value |
|---|---|
| Paper | Critique of Agent Model |
| arXiv ID | 2606.23991 |
| Submitted | June 22, 2026 (v1) |
| Authors | Eric Xing, Mingkai Deng, Jinyu Hou |
| Subjects | cs.AI, cs.LG, cs.MA, cs.RO |
| Core terms | Agentic vs agentive, GIC architecture |
| Five dimensions | Goal, identity, decision-making, self-regulation, learning |
Agentic vs agentive: the paper's central distinction
Most LLM products marketed as agents in 2026 are agentic, in the paper's vocabulary:
- Agentic — competence resides in engineered workflows. Tool definitions, permission prompts, MCP servers, subagent graphs, and Claude Code hooks live outside the model. The model reasons inside a harness someone else built.
- Agentive — capabilities arise endogenously. Goal structures, identity, regulation, and learning are internalized within the system itself, not assembled through external scaffolding.
This is not semantic nitpicking. The paper claims genuine agency requires internalization across all five dimensions. A system that only looks autonomous because a developer wrapped it in a loop is still automation with good UX.
For practitioners, the mapping is immediate:
| What you use today | Paper classification | Where competence lives |
|---|---|---|
| Claude Code + hooks + bash | Agentic | Harness, settings.json, tool policies |
| Cursor + rules + MCP | Agentic | IDE orchestration + model |
| Goal command long runs | Agentic (strong) | Prescribed goal + external loop |
| Hypothetical GIC agent | Agentive (target) | Internal goals, identity, world model |
If you have shipped with agent skills and MCP, you are building agentic systems — and that can be exactly the right engineering choice for production.
Five dimensions of agency
The paper analyzes architectures along five axes. Each must be internalized for agentive status, not delegated entirely to the environment.
1. Goal
How does the system represent and decompose objectives? Prescribed task lists and one-shot prompts are external goals. Hierarchical goal decomposition inside the agent — with sub-goals that persist and revise — is what GIC targets.
Related reading: Goal mode for long-running agents.
2. Identity
Who is the agent, and does that persist across sessions? Marketing personas and system prompts are shallow identity. Identity evolution — stable self-model that updates from experience — is the agentive bar.
3. Decision-making
Does the system choose actions from internal state, or only react to the next tool slot in a graph? Agentic stacks often hard-code decision topology. Agentive systems reason with simulative options before acting.
4. Self-regulation
Stop rules, budget caps, and human approval gates are external regulation — necessary for safety, but not the same as learned self-regulation inside the agent.
5. Learning
Fine-tuning pipelines and RAG updates happen outside the runtime loop. Self-directed learning from real and simulated experience — without a human relabeling every step — is the fifth internalized requirement.
The paper connects this framework to Descartes' grounding of agency in independent thought and to autonomous beings in science fiction — not as fluff, but as intuition pumps for what "internal" means.
GIC: Goal-Identity-Configurator architecture
Building on the critique, the authors propose GIC as a blueprint for a general-purpose agent model:
GIC stack (conceptual)
├── Hierarchical goal decomposition
├── Identity evolution (persistent self-model)
├── Simulative reasoning ← separate world model
├── Learned self-regulation
└── Self-directed learning (real + simulated experience)
Simulative reasoning is a standout detail: GIC assumes a separately trained world model the agent uses to imagine outcomes before committing to actions — closer to model-based RL than pure ReAct tool loops.
Self-directed learning spans both live environment interaction and simulation — aligning with trends in world-model agents and self-improving harnesses, but pushing capability inside the agent rather than in the harness alone.
GIC is a research architecture, not a GitHub repo you can git clone today. Treat it as a specification for where agentive research may go, not a replacement for Claude Code tomorrow morning.
What this means for coding agents in 2026
The paper does not tell you to stop using Claude Code. It tells you to name what you have:
-
Agentic systems can be excellent — Most production value in 2026 comes from agentic stacks: clear tools, permission boundaries, observable hooks, and human-in-the-loop approval. That matches how sound and traffic-light notifications keep long runs safe.
-
"Agent" marketing oversells autonomy — If every dimension is external, you have a sophisticated workflow engine, not an open-world autonomous agent. That is fine — as long as security and compliance teams know the difference.
-
Agentive research raises the oversight bar — More internal autonomy implies harder auditability. The paper explicitly discusses controllability and safety under human oversight — the same concerns driving premortem skills and enterprise agent governance.
-
Harness engineering still matters — Work like Self-Harness shows harness-only gains of 14–21 points on Terminal-Bench 2.0. Xing et al. would classify that as optimizing agentic layers — valuable, orthogonal to internalizing agency.
Safety, auditability, and human oversight
Existential "machine agency" fears assume agentive systems without oversight. The paper's closing emphasis is different: agentive systems with greater autonomy that remain under human oversight.
Practical implications:
- Auditability — Can you trace why an internal goal changed? External hooks log tool calls; internal identity evolution needs its own audit trail.
- Controllability — Kill switches and budget caps are table stakes. Learned self-regulation must not bypass them.
- Human oversight — Approval gates (red traffic lights, permission prompts) stay relevant even as architectures become more internal.
For teams shipping today, the actionable takeaway is to document which of the five dimensions are external in your stack — and which you are falsely attributing to the model.
How GIC compares to common agent patterns
| Pattern | Goal | Identity | Decision | Regulation | Learning | Paper label |
|---|---|---|---|---|---|---|
| Chatbot + RAG | External | External | External | External | External | Automation |
| ReAct tool loop | External | Prompt-only | Partial | External | External | Agentic |
| Claude Code + MCP + hooks | External | Session | Harness | Hooks + user | External | Agentic |
| Multi-agent orchestration | Shared external | Role prompts | Graph | Per-agent caps | External | Agentic |
| GIC (proposed) | Internal hierarchy | Evolving | Simulative WM | Learned | Self-directed | Agentive target |
Key quotes and framing from the abstract
The abstract states the problem plainly:
With the rise of Large Language Model (LLM) systems marketed as "coding agents", "AI co-scientists", and other "agentic" tools … it has become essential to clarify where automation ends and agency begins.
And the core technical claim:
Genuine agency requires these structures to be internalized within the system itself rather than assembled through external scaffolding.
That single sentence is the paper's contribution to every 2026 architecture review.
Related reading on explainx.ai
- Google AI Scientist / ScientistOne at ICML 2026 — "co-scientist" marketing vs end-to-end research harness
- What Are AI Agents? Complete guide for 2026 — agent loop, tools, memory
- Self-Harness: agents that improve their own framework — harness-side optimization
- Goal mode and long-running agents — external goal persistence
- Why AI companies want you using agents — economics of the agent label
- Steering Claude Code: CLAUDE.md, skills, hooks — agentic scaffolding in practice
Primary source: arXiv:2606.23991 — Critique of Agent Model · DOI 10.48550/arXiv.2606.23991
Paper details, author list, and arXiv metadata accurate as of June 23, 2026. GIC is a proposed research architecture — verify against the PDF before citing implementation claims in production docs.