Source-linked AI summary
Critique of Agent Model
Eric Xing, Mingkai Deng, Jinyu Hou
TL;DR
Current AI agents often achieve complex tasks through external scaffolding, leaving agency insufficiently defined. The paper distinguishes agentic from agentive systems and proposes GIC, which internalizes goal, identity, reasoning, self-regulation, and learning within an adaptive architecture.
Problem
Many systems labeled agents rely on externally specified tools and workflows, while learning processes remain external, limiting analysis of general-purpose agency.
Method
The paper surveys agent architectures across five dimensions and proposes GIC, combining hierarchical goals, evolving identity, world-model simulation, learned self-regulation, and self-directed learning.
Results
The analysis distinguishes agentic systems, whose competence resides in engineered workflows, from agentive systems, whose capabilities arise endogenously within a single adaptive system.
Takeaways & Limitations
Genuine agency is framed as internalizing goal-oriented action, adaptive identity, self-regulated deliberation, autonomous learning, and social participation rather than accumulating external scaffolding.
Takeaways & Limitations
GIC training may still produce mistakes, including harmful ones, even though simulation, oversight, and modularity are proposed to manage these risks.
Abstract
from arXiv · showhide
What is an agent? What constitutes agency? With the rise of Large Language Model (LLM) systems marketed as ``coding agents'', ``AI co-scientists'', and other ``agentic" tools that promise to drive up productivity, and at the same time, ``existential" concerns such as AI escaping human control with destructive power under a speculative ``machine agency" against humans, it has become essential to clarify where automation ends and agency begins, both for building capable systems and for understanding whether and what to fear. Drawing on Descartes' grounding of agency in independent thought, and on portrayals of autonomous beings in science fiction, we survey the current landscape of AI agents, and analyze agent architectures along five dimensions: goal, identity, decision-making, self-regulation, and learning. Specifically, we argue that genuine agency requires these structures to be \emph{internalized within the system itself} rather than assembled through external scaffolding. This distinction between \emph{agentic} systems, whose competence resides in engineered workflows, and \emph{agentive} systems, whose capabilities (including social interaction) arise endogenously, defines the boundary between systems designed for prescribed tasks, and those capable of operating in the open world with true autonomy. Building on this analysis, we propose the Goal-Identity-Configurator (GIC) architecture for a general-purpose agent model, combining hierarchical goal decomposition, identity evolution, simulative reasoning grounded in a separately trained world model, learned self-regulation, and self-directed learning from both real and simulated experience. Furthermore, we share insight on the auditability, controllability, and safety of agentive systems that possess greater autonomy and ``agency", but remain under human oversight.
1 Introduction
The paper distinguishes agentic systems, whose competence depends on externally engineered tools and workflows, from agentive systems, whose agency emerges endogenously through internal goals, identity, decision-making, self-regulation, and learning. It proposes the Goal-Identity-Configurator architecture, pairing an agent model that decides what to do with a separately learned world model that predicts what will happen.
- Agency and autonomy: Genuine agency involves long-term goals, evolving identity, purposeful planning, social relationships, self-regulation, and self-reflection.These properties are associated with agency across philosophical, psychological, sociological, and economic traditions.
- Agentic versus agentive systems: Agentic systems complete complex tasks through externally orchestrated tools, workflows, and control loops, whereas agentive systems derive capabilities endogenously.Agentic competence resides primarily in engineering around a reasoning model such as an LLM; agentive systems maintain goals and operate with fuller agency.
- Agentic versus agentive systems: The paper argues that agency should emerge from a model that develops identity, pursues goals, and organizes behavior across diverse environments rather than accumulating external scaffolding.The motivation is that enumerating every behavior through tools, prompts, or skills may not scale to biological agents’ diversity and adaptability across linguistic, physical, spatial, and social capabilities.
- Agent model: Conditioning an agent model on goal g and identity i enables it to inspect, decompose, and revise long-term objectives and its self-model.The model maps world state s to predicted action a while supporting abstract reasoning and physical-world action through end actuators.
- Goal-Identity-Configurator architecture: The GIC architecture combines hierarchical persistent goals, evolving identity, world-model simulation, learned self-regulation, and self-directed learning from real and simulated experience.Its agent model remains functionally distinct from a separately learned world model: the former decides what to do, while the latter predicts what will happen.
2 The Boundary Between Agentic and Agentive Systems
Agentic systems complete tasks through externally orchestrated tools and workflows, whereas agentive systems derive capabilities from internal organization. Genuine agency therefore depends on internally organized goals, identity, decision-making, self-regulation, and learning that can adapt and revise behavior.
- Decision-making: An agent model decides actions, while a distinct world model predicts environmental transitions and supports simulated evaluation of future action consequences.The agent factor determines what to do, whereas the universe factor determines what happens next; the world model approximates the latter.
- Goal: Agentic systems execute externally specified instructions, while agentive systems maintain, decompose, and revise their own goals during ongoing decision-making.The degree of endogenous goal formation, decomposition, and maintenance marks one axis from agentic to agentive systems.
- Identity: Agentive identity is an evolving internal state that updates capabilities, role assumptions, affordances, and relationships in response to environmental feedback.This contrasts with identity fixed by system design, such as system prompts or configuration files.
- Decision-making: Agentive decision-making includes adapting internal computation to urgency, difficulty, uncertainty, and resource constraints rather than applying one fixed deliberation strategy.Decisions may range from direct policy execution to extended deliberation or replanning, with reactive execution handling fine-grained actions.
- Learning: Learning differentiates systems whose behavior is fixed before deployment from systems that can improve through never-ending learning, recursive self-improvement, or auto research.Existing systems commonly require external retraining or prompt redesign for post-deployment behavioral change.
- Boundary: Across these dimensions, genuine agency arises from an internal configurator that adapts, revises, and organizes decision-making processes instead of externally engineered pipelines prescribing behavior.The relevant processes include planning, self-regulation, learning, and interaction.
3 Landscape of Systems Labeled as “Agents”
Systems labeled “agents” span deterministic scripts, LLM wrappers, model-centered systems, and embodied platforms, but differ in where behavioral organization resides. The landscape increasingly internalizes planning, acting, and adaptation, while much current capability still depends on external orchestration, domain restriction, and engineered workflows.
- Landscape overview: The agent label covers systems from simple automation scripts to embodied learning systems, obscuring differences in how behavior is organized.The landscape is examined by behavior-producing mechanisms rather than application domains.
- Program-Based Systems and Classical Bots: Program-based systems and classical bots pursue explicit goals through deterministic logic, fixed control rules, pattern matching, or scripted interaction sequences.Examples include thermostats, ELIZA, Selenium, and Playwright.
- LLM Wrapper Systems: LLM wrapper systems place pretrained language models inside orchestration layers that prescribe behavior through planning loops, tool pipelines, or multi-agent coordination graphs.Examples include DeerFlow, Agent Skills, AutoGen, Decagon, and Cursor.
- LLM-Centered Systems: LLM-centered systems shift more behavioral structure into models trained or fine-tuned to map observations to actions over extended trajectories.Applications include browser use, deep research, software engineering, and game playing.
- Model-less Physical Systems: Physical embodiment does not by itself establish internally organized decision-making: robots may combine learned low-level control with externally scripted task logic.Industrial robots execute programmed routines, while legged platforms often pair learned control with scripted tasks.
- Landscape overview: Overall, much recent capability derives from external orchestration, narrow domains, and powerful foundation models within engineered workflows, leaving core agency structures outside the model in many cases.These structures include endogenous goal decomposition, persistent self-models, adaptive self-regulation, continual learning, and autonomous social organization.
4 Critique of Agent Modeling
The critique argues that externally supplied goals, fixed identities, and ungrounded computation limit scalable, general-purpose agency. It proposes internal goal decomposition, identity revision, and selectively invoked world-model simulation as foundations for more autonomous systems.
- Goal: Externally supplied, short-horizon goals make systems controllable for narrow tasks but difficult to scale toward long-horizon autonomy.Their objectives disappear when interactions end, whereas autonomous agents should pursue extended goals without step-by-step human guidance.
- Goal: A learned goal-decomposition module δ can break long-term goals into dependency- and priority-ordered subgoals that remain revisable as information changes.This isolates long-term planning in δ while allowing individual subgoals to be executed by short-horizon mechanisms.
- Identity: Fixed, externally specified identities cannot adapt to unexpected capabilities, limitations, environments, or feedback, motivating autonomous revision of the agent’s self-model.The proposed fast-slow mechanism combines stepwise self-model updates with periodic retraining, improving decisions immediately and supplying higher-quality data for later updates.
- Decision-making: Internal computation or reasoning tokens alone do not provide grounded counterfactual planning; an explicit world model supplies the missing action-consequence engine.The mixed policy πmix selectively uses world-model simulation and otherwise falls back to the baseline, guaranteeing at least equal value and strictly better performance when an improvement is adopted.
- Self-regulation: Fixed planning pipelines and pure H-step MPC are inadequate as universal decision procedures because precision requires increasingly costly simulation depth.The agent should govern when to use reactive action versus simulative reasoning according to urgency, difficulty, uncertainty, and resource constraints.
5 The GIC Agent Model
The GIC Agent Model integrates persistent hierarchical goals, evolving identity, simulative reasoning, learned self-regulation, and self-directed learning within a single learned system. It supports persistent operation, PEG evaluation, and layered auditability through explicit, interpretable capabilities and human oversight.
- Architecture: GIC combines hierarchical goal decomposition, evolving identity, internal-world-model reasoning, learned self-regulation, and self-directed learning in a single learned model.These requirements are presented as necessary for capability akin to genuine agency.
- Aircraft-pilot use case: The aircraft-pilot use case demonstrates staged development through ground-school knowledge, simulator practice, real-world adaptation, and fleet coordination.The stages develop world modeling, reactive and deliberate competence, identity calibration, online adaptation, and social coordination.
- Persistent operation: GIC supports persistent operation with minimal external intervention by integrating planning, acting, reflection, training, and self-model adjustment into continuous operation.Inference and learning are described as a single process rather than separate phases.
- Evaluation: PEG evaluates agentive systems through Performance, Efficiency, and Growth, covering goal decomposition, identity evolution, simulative reasoning, reactive execution, self-regulation, and self-directed learning.Preliminary results provide initial evidence for Performance and Efficiency, while Growth evaluation remains incomplete.
- Safety and oversight: GIC’s explicit modules enable layered auditability, human oversight, targeted diagnosis, and correction of undesirable subgoals before execution.Training primarily occurs in a world-model sandbox, helping confine mistakes before real deployment.
6 Conclusion
The conclusion argues that genuine agency requires internalizing core agent characteristics within a standalone system rather than relying on external scaffolding. It presents GIC as a constructive alternative and outlines future directions for broader, open-ended agent modeling.
- Core argument: The authors frame the paper as an invitation to deeper reflection on what agents are, what constitutes genuine agency, and how to build useful agent models.They explicitly present these as fundamental questions rather than definitive answers.
- Core argument: Genuine agency requires a standalone system to internalize goal-oriented action, adaptive identity, self-regulated deliberation, autonomous learning, and emergent social participation.The authors contrast this with externally orchestrated tools and workflows, which characterize agentic systems; current efforts remain primitive.
- GIC architecture: The GIC architecture combines goal decomposition, identity evolution, simulative reasoning, self-regulation, and self-directed learning with a separately learned world model.The proposal is intended to support further advances in both theory and implementations of stronger agent models.
- Future directions: Future directions include multi-agent modeling, interactions across different time scales and modalities, and autonomous perpetual learning in open-ended environments.The envisioned applications include modeling collective behaviors of businesses or societies and consequences for public health.
A Detailed Restatement and Proof for Theorem 1
Theorem 1 shows that fast-slow learning achieves at least as low cumulative regret as slow-only learning when identity revision improves decisions and slow updates are jointly monotone. The proof decomposes the advantage into within-round gains and cross-round carry-over, with the latter preserved but not necessarily amplified by slow updates.
- Theorem statement: Under A1 and A2, fast-slow learning dominates slow-only learning through nonnegative per-round advantage ηk ≥ 0.The two base-policy sequences can diverge after round 1, so the result requires induction rather than a single monotonicity application.
- Step 1: Per-step value improvement: Identity revision yields a strictly positive per-step gain when (1 − δ)λ > δB, with δ = δ1 + δ2 − δ1δ2 < 1.The gain follows because the favorable event occurs with probability at least (1 − δ1)(1 − δ2), while complementary outcomes incur degradation bounded by B.
- Step 3: Cross-round compounding: Joint monotonicity carries the fast-slow advantage across rounds because it begins each update with both a stronger base policy and stronger behavioral-policy experience.The induction establishes ηk+1 ≥ 0, preserving the advantage through every subsequent slow update.
- Combining the terms: The resulting advantage increases with total interaction steps and is non-decreasing in the number of update cycles.Slow updates cannot erase fast adaptation’s gains, but A2 does not require them to amplify those gains; monotonicity can be tested for specific update operators.
- Step 2: Within-round regret reduction: Summed across interactions, identity revision contributes a within-round gain of PK k=1 Nk ε̄, even without further slow updates.This is the first component of the regret reduction; the remaining component comes from cross-round differences in base-policy quality.
B Proof for Theorem 2
The proof defines a mixed decision rule that uses the world-model policy only when it appears clearly better than the base policy, allowing for model error. It then combines value and advantage-function bounds with the Simulation and Performance Difference Lemmas to establish the claimed performance relation.
- Value comparison: The proof compares state and action values under the true environment µ and world model f using the Simulation Lemma.The corresponding quantities are defined for a fixed policy π, and their differences are bounded for every state-action pair.
- Advantage comparison: The proof introduces advantage functions to measure how much better an action is than continuing with π in the true environment and world model.The true-environment advantage is defined explicitly, with an analogous definition under f.
- Mixed decision rule: The mixed rule πmix = ϕ(π, f, ϵ) follows the world-model policy only when its predicted advantage exceeds the base policy by 2ϵmodel.Otherwise, it retains π, leaving a margin 2ϵmodel for model error.
- Performance guarantee: Applying the Performance Difference Lemma bounds the mixed policy’s value using the marginal state distribution dπmix µ induced in the true environment.The resulting inequality is strict whenever πmix adopts π∗f on a set of states having nonzero probability under dπmix µ.
C Proof for Theorem 3
The proof establishes Theorem 3 through four steps using an augmented reward derived from the cost function, Bellman-operator contraction, finite-horizon MPC interpretation, and reward alignment. It concludes that the MPC policy matches the optimal policy and derives the stated bound when γ and Cmax are constants.
- Step 1: The proof defines the augmented reward as ˜r(s, g) = −Cg(s) and uses the corresponding Bellman operators, whose value functions are unique fixed points.The operators are introduced for both the augmented optimality operator and any policy-specific operator.
- Step 1: For a policy greedy with respect to any bounded value function V, the proof establishes the Step 1 inequality using Equation 19, sup norms, and the γ-contraction property.The argument proceeds through contraction and triangle-inequality bounds before proving the claim.
- Step 2: The iterate ˆV (H−1) = ˜T H−10 represents the augmented reward of a finite-horizon rollout with zero terminal value, so pure H-step MPC is greedy with respect to it.Applying the γ-contraction of ˜T then relates this finite-horizon value to the relevant optimal value.
- Step 4: Because ˜r(s, g) = −Cg(s) = r(s, g) −bg, maximizing ˜Vπ,f is equivalent to maximizing V g π,f, yielding ˜π∗ = π∗.The difference is the policy-independent constant bg 1−γ.
D Proof for Theorem 4
The proof compares a policy’s value in mixed experience with its value in the real environment, then uses computational inclusion and chained inequalities to establish the theorem.
- Proof structure: The Simulation Lemma and mixed-experience definition determine the difference between a policy’s values in Mα and the real environment µ.The passage introduces the value comparison for an arbitrary policy π.
- Proof structure: By construction, Πenv(Dµ) is contained in Πmix(Dµ, Df), since access to world model f and extra simulated experience cannot reduce computability.This establishes the relevant inclusion between the environment and mixed-experience policy classes.
- Proof structure: Inequality 27 supplies the next bound used in the proof.The passage states that the inequality is invoked without reproducing its contents.
- Proof structure: The proof concludes by chaining the established inequalities.The final passage identifies the concluding proof operation but does not display the resulting expression.