From Language Models to World-Acting Systems: Progress and Limits of Agentic AI across Digital, Social, Virtual, and Physical Environments

2609.04894 ↗ · 2026-09-04 · cs.AI, cs.LG, cs.MA

Linsen Zhu, Mengqing Cai

Review article. 29 pages, 1 figure, 3 tables. Literature cutoff: 31 August 2026

The finding

Headline claim

Within the public evidence through August 2026, agentic AI systems' action interfaces and deployment scope have expanded faster than evidence for verifiable autonomy: reliable completion, recovery, authorization, and independent verification remain the weak links, so delegation should be expanded only where bounded evidence supports it.

The test

A critical (not systematic or meta-analytic) review of representative benchmarks, peer-reviewed experiments, protocol specifications (MCP, A2A), preprints, and first-party system cards, organized by delegated authority, temporal persistence, and environmental coupling, with a model-harness-environment-delegator unit of analysis.

Effect size & scope

Anchor numbers come from the papers' cited benchmarks: original WebArena best GPT-4 agent 14.41% vs 78.24% human; original OSWorld best below 12.2% vs 72.4% human; SWE-agent 12.5% pass@1; tau-bench shows low pass^k reliability even when pass@1 is acceptable. Later first-party computer-use systems report large gains but still concede oversight needs. No pooled effect size is offered, and the review itself warns scores are configuration-specific and quickly stale.

Limitations

The authors concede this is a critical review with no corpus-wide prevalence estimates, no meta-analytic effect sizes, representative rather than exhaustive sampling, and category assignments frozen at the cutoff. Unstated: the core asymmetry claim is an interpretive judgment, not a measurement, and the anchor benchmarks are two-plus years old, so the gap may have narrowed substantially.

Counterpoint

A skeptic would say the conclusion is close to definitional: verification always lags capability in any fast-moving field, and citing 2023-era WebArena and OSWorld numbers in a 2026 review stacks the deck. First-party agent products already complete real work at scale, and 'verifiable autonomy' is a bar no deployed software system of any kind clears, so the framing may be a safety literature template rather than an empirical finding.

Economy implication

Little or nothing quantitative yet for the AI economy: the paper offers no cost, revenue, or adoption data. Its indirect implication is that enterprise agentic ROI will be gated by verification, oversight, and permission infrastructure, not model capability alone, which supports near-term spending on harness and governance tooling but is asserted, not measured.

Who this matters for

Technical Capability: Practitioners should stop reading model benchmark scores as system readiness: the review shows harness and interface design changes outcomes at fixed weights, so evals must be run on the configured system.

Regulatory & Legal: The justified-delegation framework and proposed reporting standard give regulators a concrete vocabulary (authority, recovery, verification) for grading agent deployments rather than banning or ignoring them.

Labor & Knowledge Work: If enterprise adoption is gated on verification and human oversight capacity, oversight becomes a scarce labor input that constrains how fast agentic automation of knowledge work can actually scale.

Bears on (advisory · papers never write evidence)

[1.3] Agentic reliability (multi-step completion without human correction) is improving materially, not just peak single-shot.

[B1] Agentic reliability gains will translate into enterprise ROI.

Abstract

Large language models become consequential agents when surrounding systems let outputs change external state. Models now call tools, operate interfaces, delegate work, retain state, inhabit generated worlds, and control robots or laboratory equipment. Such advances are often narrated as one march toward autonomy, conflating model competence, system integration, persistence, and safe authority. This critical review synthesizes primary research and official technical specifications available by 31 August 2026. We organize the evidence along delegated authority, temporal persistence, and environmental coupling, while separating model, harness, and environment. Within the evidence examined, action-interface expansion is documented more convincingly than robust completion, recovery, authorization, or independent verification. Model Context Protocol and Agent2Agent improve interoperability but do not establish trustworthy delegation; multi-agent organization adds specialization alongside cost and correlated failure. Persistent simulations and world models support training and planning but do not themselves demonstrate agency; robotics and self-driving laboratories establish bounded feasibility rather than unattended open-world reliability. We propose justified delegation as an analytical and normative heuristic, not an observed law or certified score: expand action scope only where evidence supports provenance, bounded authority, failure detection, safe recovery, and calibrated human control. This framing yields a research agenda for coupled model-harness evaluation, capability-based permissions, durable state, cross-agent accountability, and staged physical validation.