arXivPaperNeeds Review

A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AI

This survey organizes long-horizon agent weaknesses into five capability gaps, offering a useful checklist for harness design and evaluation rather than a new implementation.

arXiv · Aug 3, 2026

Open Source Open Markdown Open JSON

Source Summary

The survey groups cognitive capability gaps into **five dimensions**: persistent state, goal-directed autonomy, self-monitoring and control, environment interaction, and learning and adaptation. It also proposes a conceptual **ACIA architecture** and discusses cognition-focused evaluation.

Practical Implication

Use the taxonomy as a review checklist for long-running coding agents: inspect how state persists, goals remain bounded, failures are detected, tools affect the environment, and feedback changes later behavior. It can also help separate harness problems from model limitations.

Agent-Ready Context

The survey groups cognitive capability gaps into **five dimensions**: persistent state, goal-directed autonomy, self-monitoring and control, environment interaction, and learning and adaptation. It also proposes a conceptual **ACIA architecture** and discusses cognition-focused evaluation.

Use the taxonomy as a review checklist for long-running coding agents: inspect how state persists, goals remain bounded, failures are detected, tools affect the environment, and feedback changes later behavior. It can also help separate harness problems from model limitations.

This is a literature-organizing framework, not a validated agent stack or benchmark result. The supplied material offers no implementation details, comparative measurements, or evidence that ACIA improves reliability in deployed systems.

Connected Context · Feed7 Judgment

This consolidates scattered agent-reliability concerns into five review dimensions, helping teams distinguish missing model capability from missing harness support. It confirms that durable state, bounded goals, monitoring, controlled action, and adaptation should be assessed separately. Because ACIA is conceptual, the taxonomy organizes requirements but does not validate an architecture or replace task-specific controls and evaluations.

AI Agents for Performance: Ship Faster, Pay Less — Rajat Shah, NetflixNetflix’s bounded optimization workflow supplies a concrete instance of the taxonomy’s state, environment-interaction, and self-monitoring dimensions through commit-linked evidence, canaries, and engineer release gates.The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic DistillationThe finding that long-horizon planning needs explicit state transitions and compositional trajectories provides experimental support for treating persistent state and goal-directed autonomy as separable capability gaps.Learning on the Job: The Future of Post-Training — Raymond Feng, Applied ComputeProduction-harness training directly concerns the taxonomy’s learning-and-adaptation dimension, while its non-replayable feedback problem shows why naming that capability does not establish a workable update mechanism.Build for the Memo, Not the Demo — Shawn Chan, China Resources HoldingsClaim provenance, uncertainty, contradiction checks, and logged approval turn the taxonomy’s self-monitoring and controlled interaction dimensions into explicit output and release contracts for a high-stakes domain.

Context Map

agentresearch#agent-memory#agent-reliability#harness-engineering

Uncertainty

This is a literature-organizing framework, not a validated agent stack or benchmark result. The supplied material offers no implementation details, comparative measurements, or evidence that ACIA improves reliability in deployed systems.