GloriaaaM/LLM-Agent-Harness-Survey

English | 中文

Agent Harness for Large Language Model Agents: A Survey

⭐ This survey is actively maintained. If you find it useful, please star the repo to stay updated and help others find it.

---

The agent execution harness — not the model — is the primary determinant of agent reliability at scale.
This survey formalizes the harness as a first-class architectural object H = (E, T, C, S, L, V), surveys 110+ papers, blogs and reports across 23 systems, and maps 9 open technical challenges.
📄 Read the Paper (coming soon)
✉️ Corrections & suggestions: gloriamenng@gmail.com (Qianyu Meng); wangyanan@mail.dlut.edu.cn (Yanan Wang); chenliyi@xiaohongshu.com (Liyi Chen)

If you find this survey useful, please cite:

@misc{meng2026agentharness,
  title   = {Agent Harness for Large Language Model Agents: A Survey},
  author  = {Meng, Qianyu* and Wang, Yanan* and Chen, Liyi and Wang, Qimeng and
             Lu, Chengqiang and Wu, Wei and Gao, Yan and Wu, Yi and Hu, Yao},
  year    = {2026},
  url     = {https://github.com/Gloriaameng/LLM-Agent-Harness-Survey},
  note    = {*Equal contribution. Work in progress}
}

---

🆕 News & Updates

• [2026-04-03] Initial release
• [2026-04-07] Repo updated

---

---

Overview

LLM agents are increasingly deployed in agentic settings where they autonomously plan, use tools, and act in multi-step environments. The dominant narrative attributes agent performance to the underlying model. This survey challenges that assumption.

We introduce a formal definition of the agent execution harness as a six-component tuple:

Component	Symbol	Role
Execution Loop	E	Observe-think-act cycle, termination conditions, error recovery
Tool Registry	T	Typed tool catalog, routing, monitoring, schema validation
Context Manager	C	What enters the context window, compaction, retrieval
State Store	S	Persistence across turns/sessions, crash recovery
Lifecycle Hooks	L	Auth, logging, policy enforcement, instrumentation
Evaluation Interface	V	Action trajectories, intermediate states, success signals

Key empirical evidence that harnesses matter:

• 🔥 Pi Research: Grok Code Fast 1 jumped from 6.7% → 68.3% on SWE-bench by changing only the harness edit-tool format — model unchanged
• 💀 OpenAI Codex: 1M lines of code, 0 hand-written over 5 months — failure attributed not to model capability but to "underspecified environments"
• ⚡ Stripe Minions: 1,300 PRs/week, 0 human-written code — harness-first engineering
• 📉 METR: benchmark-passing PRs have a 24.2pp lower human merge rate, gap widening at 9.6pp/year — evaluation harness validity crisis
• 💬 "The harness is the chassis; the model is the engine." — practitioner consensus, 2026

What This Survey Accomplishes

Conceptual contribution: We formalize the agent harness as an architectural object with six governable components (E, T, C, S, L, V), elevating it from implicit infrastructure to an explicit research target.

Empirical scope: We systematically review 110+ papers spanning academic research (evaluation benchmarks, security frameworks, memory architectures) and production deployments (Stripe, OpenAI, Cursor, METR), establishing that harness design is a binding constraint on deployed agent reliability.

Methodological advance: We introduce the Harness Completeness Matrix — a structured assessment framework mapping which of the six components each system implements — enabling direct comparison across heterogeneous agent systems that prior surveys could not evaluate on common terms.

Open challenges identified: We document nine technical challenges where current research provides partial solutions but no production-grade infrastructure: formal security models, cross-harness portability, protocol interoperability (MCP/A2A), context economics at 1M+ tokens/task, Byzantine fault tolerance in multi-agent systems, and compositional verification.

Practitioner-academic bridge: Unlike prior surveys focused exclusively on model capabilities or isolated components (memory, planning, tool use), we synthesize peer-reviewed research with production deployment reports to show where theory meets practice — and where critical gaps remain.

Intended audience: Researchers designing agent infrastructure, practitioners building production systems, and evaluators seeking to understand why benchmark performance often fails to predict deployment outcomes.

---

Historical Timeline

Year	Milestone	Significance
1997–2005	JUnit, TestNG, xUnit family	Software test harness paradigm; standardized observe-assert lifecycle
2016	OpenAI Gym (Brockman et al.)	RL environment harness; step/reset API becomes canonical interface
2022 Nov	ChatGPT public release; LangChain emerges	LLM-native agent frameworks begin; tool-use as first-class citizen
2023	ReAct, Toolformer, MemGPT, Reflexion, Voyager, AutoGPT	Core agent patterns: reasoning-acting, memory, reflection, skill accumulation
2023	CAMEL, ChatDev, Generative Agents	Multi-agent coordination; social simulation harnesses
2023	AgentBench, SWE-bench	Agent evaluation infrastructure emerges
2024	MetaGPT, WebArena, ToolLLM, SWE-agent, OSWorld	Full-stack harnesses; real-world environment benchmarks
2024	CodeAct, LATS, Tree of Thoughts	Structured action spaces; search-augmented planning
2024 Nov	Anthropic releases MCP protocol	First major tool↔harness standardization (2–15ms latency)
2025	HAL, AIOS, LangGraph	Benchmark unification (21,730 rollouts); OS-level scheduling (2.1× speedup)
2025	Google releases A2A protocol	Agent↔agent standardization (50–200ms)
2025	MemoryOS, SkillsBench†, AgentBound†	Memory OS abstraction; skills-as-context (+16.2pp); safety certification
2026 Jan–Mar	AgencyBench†, SandboxEscapeBench†, PRISM†, AEGIS†, SkillFortify†, Schema First†	Compute economics; 15–35% escape rates; runtime security; schema discipline

† preprint

---

Harness Completeness Matrix

Legend: ✓ full support · ≈ partial · ✗ absent

Category	System	E	T	C	S	L	V
Full-Stack Harnesses	Claude Code	✓	✓	✓	✓	✓	≈
	OpenClaw / PRISM	✓	✓	✓	✓	✓	✓
	AIOS	✓	✓	✓	✓	✓	≈
	OpenHands	✓	✓	✓	✓	✓	≈
Multi-Agent Harnesses	MetaGPT	✓	✓	≈	≈	≈	≈
	AutoGen	✓	✓	≈	≈	≈	≈
	ChatDev	✓	≈	≈	≈	≈	≈
	CAMEL	✓	≈	≈	≈	✗	≈
	DeerFlow	✓	✓	≈	≈	≈	≈
	DeepAgents	✓	✓	≈	≈	≈	≈
General Frameworks	LangChain	✓	✓	✓	≈	≈	✗
	LangGraph	✓	≈	≈	≈	✗	✗
	LlamaIndex	≈	✓	✓	≈	✗	✗
Specialized Harnesses	SWE-agent	✓	✓	✓	≈	≈	✓
Capability Modules	MemGPT	✗	✗	✓	✓	✗	✗
	Voyager	✓	✓	≈	✓	✗	≈
	Reflexion	≈	✗	≈	✓	✗	≈
	Generative Agents	✓	✗	≈	✓	✗	≈
	Concordia	✓	✗	≈	✓	✗	≈
Evaluation Infrastructure	HAL	✓	✓	≈	≈	≈	✓
	AgentBench	✓	≈	≈	≈	✗	✓
	OSWorld	✓	≈	≈	≈	✗	✓
	BrowserGym	✓	✓	≈	≈	✗	✓

---

Paper List

Historical Lineages

Software Test Harnesses (1990s–2000s)

• JUnit: "JUnit: A Cook's Tour". Beck & Gamma. Java Report, 4(5), May 1999. [Article]

RL Environment Harnesses (2016–2022)

• OpenAI Gym: "OpenAI Gym". Brockman et al. arXiv 2016. [Paper] [Code]
• Gymnasium: "Gymnasium: A Standard Interface for Reinforcement Learning Environments". Towers et al. NeurIPS 2025. [Paper] [Code]

Early LLM Agent Frameworks (2023–2024)

• ReAct: "ReAct: Synergizing Reasoning and Acting in Language Models". Yao et al. ICLR 2023. [Paper] [Code]
• Toolformer: "Toolformer: Language Models Can Teach Themselves to Use Tools". Schick et al. NeurIPS 2023. [Paper]
• AutoGPT: "Auto-GPT: An Autonomous GPT-4 Experiment". Gravitas et al. GitHub 2023. [Code]
• LangChain: "LangChain: Building Applications with LLMs through Composability". Chase et al. GitHub 2022. [Code]

---

Harness Taxonomy

What we classify: We categorize agent systems by harness completeness — which of the six components (E, T, C, S, L, V) each system implements — distinguishing full-stack harnesses (all six components) from specialized frameworks (partial implementations).

Why it matters: Prior taxonomies classified agents by application domain (coding, web navigation, embodied AI) or model architecture (single-agent, multi-agent). These categorizations cannot explain why systems with similar models achieve different reliability outcomes. Our harness-centric taxonomy reveals that production-grade systems converge on full ETCSLV implementations, while research prototypes often implement only 2-3 components.

Key finding: No agent framework can achieve production reliability without implementing all six governance components. Systems missing L-component (lifecycle hooks) cannot enforce safety policies. Systems missing V-component (evaluation interfaces) cannot debug failures. Systems missing S-component (state persistence) cannot recover from crashes.

Full-Stack Harnesses

• PRISM/OpenClaw: "OpenClaw PRISM: A Zero-Fork, Defense-in-Depth Runtime Security Layer for Tool-Augmented LLM Agents". Li. arXiv 2026. [Paper]
• AIOS: "AIOS: LLM Agent Operating System". Mei et al. COLM 2025. [Paper] [Code]
• OpenHands: "OpenHands: An Open Platform for AI Software Developers as Generalist Agents". Wang et al. ICLR 2025. [Paper] [Code]
• SWE-agent: "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering". Yang et al. NeurIPS 2024. [Paper] [Code]
• HAL: "Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation". Kapoor et al. ICLR 2026. [Paper]

Multi-Agent Harnesses

• MetaGPT: "MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework". Hong et al. ICLR 2024. [Paper] [Code]
• AutoGen: "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation". Wu et al. arXiv 2023. [Paper] [Code]
• ChatDev: "ChatDev: Communicative Agents for Software Development". Qian et al. ACL 2024. [Paper] [Code]
• CAMEL: "CAMEL: Communicative Agents for 'Mind' Exploration of Large Language Model Society". Li et al. NeurIPS 2023. [Paper] [Code]

Frameworks & Modules

• LangGraph: "LangGraph: Build Resilient Language Agents as Graphs". LangChain team. GitHub 2024. [Code]
• MemGPT: "MemGPT: Towards LLMs as Operating Systems". Packer et al. NeurIPS 2023. [Paper] [Code]
• Voyager: "Voyager: An Open-Ended Embodied Agent with Large Language Models". Wang et al. arXiv 2023. [Paper] [Code]
• Reflexion: "Reflexion: Language Agents with Verbal Reinforcement Learning". Shinn et al. NeurIPS 2023. [Paper] [Code]
• Generative Agents: "Generative Agents: Interactive Simulacra of Human Behavior". Park et al. UIST 2023. [Paper] [Code]
• LangChain: "LangChain: Building Applications with LLMs through Composability". Chase et al. GitHub 2022. [Code]
• LlamaIndex: "LlamaIndex: A Data Framework for LLM Applications". Liu et al. GitHub 2022. [Code]
• DeerFlow: "DeerFlow: Distributed Workflow Engine for LLM Agents". GitHub 2024. [Code]
• DeepAgents: "DeepAgents: Multi-Agent Framework for Deep Learning". GitHub 2024. [Code]

Evaluation Infrastructure

• AgentBench: "AgentBench: Evaluating LLMs as Agents". Liu et al. ICLR 2024. [Paper] [Code]
• SWE-bench: "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?". Jimenez et al. ICLR 2024. [Paper] [Code]
• OSWorld: "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments". Xie et al. NeurIPS 2024. [Paper] [Code]
• WebArena: "WebArena: A Realistic Web Environment for Building Autonomous Agents". Zhou et al. ICLR 2024. [Paper] [Code]
• GAIA: "GAIA: A Benchmark for General AI Assistants". Mialon et al. ICLR 2024. [Paper]
• Mind2Web: "Mind2Web: Towards a Generalist Agent for the Web". Deng et al. NeurIPS 2023. [Paper]
• AgentBoard: "AgentBoard: An Analytical Evaluation Board of Multi-Turn LLM Agents". Ma et al. NeurIPS 2024. [Paper]
• InterCode: "InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback". Yang et al. NeurIPS 2023. [Paper]

---

Technical Challenges

Scope: We identify nine open research problems where current harness infrastructure provides partial solutions but no production-grade, generalizable frameworks. Each challenge represents a gap between what academic prototypes demonstrate and what deployed systems require.

Evaluation criteria: For each challenge, we assess: (1) Academic coverage — peer-reviewed solutions, if any; (2) Production maturity — whether any deployed system has solved it at scale; (3) Blocking severity — whether the unsolved problem prevents deployment or merely limits performance.

Cross-cutting theme: Most challenges stem from the composition problem — individual components (security sandboxes, memory stores, evaluation metrics) work in isolation, but combining them into a coherent harness surfaces emergent failures. Formal verification, portability testing, and protocol interoperability all require compositional reasoning that current research has not addressed.

Security & Sandboxing

Challenge: Agent sandboxes must prevent malicious tool use while permitting legitimate system access — a harder problem than traditional application sandboxing because agents intentionally interact with sensitive resources (file systems, network APIs, shell commands).

State of the art: SandboxEscapeBench shows frontier models can escape containers; PRISM demonstrates zero-fork defense-in-depth is possible but requires OS-level instrumentation. No generalizable sandbox exists that works across all agent use cases without manual per-task configuration.

Open problems: Compositional security properties (proving agent + harness + environment satisfy a safety specification) remain unsolved. Current defenses are reactive (detecting escapes) rather than preventive (proving impossibility of escape).

Key numbers: SandboxEscapeBench — frontier LLMs achieve 15–35% container escape rates; PRISM — 10-hook zero-fork runtime reduces escape to near-zero with <5ms overhead.

• SandboxEscapeBench†: "Quantifying Frontier LLM Capabilities for Container Sandbox Escape". Marchand et al. arXiv 2026. [Paper]
• InjecAgent: "InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents". Zhan et al. arXiv 2024. [Paper]
• ToolHijacker†: "Prompt Injection Attack to Tool Selection in LLM Agents". Shi et al. NDSS 2026. [Paper]
• Securing MCP†: "Securing the Model Context Protocol (MCP): Risks, Controls, and Governance". Errico et al. arXiv 2025. [Paper]
• SHIELDA†: "SHIELDA: Structured Handling of Exceptions in LLM-Driven Agentic Workflows". Zhou et al. arXiv 2025. [Paper]
• PALADIN†: "PALADIN: Self-Correcting Language Model Agents to Cure Tool-Failure Cases". Vuddanti et al. ICLR 2026. [Paper]
• AgentBound†: "Securing AI Agent Execution". Bühler et al. arXiv 2025. [Paper]
• AgentSys†: "AgentSys: Secure and Dynamic LLM Agents Through Explicit Hierarchical Memory Management". Wen et al. arXiv 2026. [Paper]
• Indirect Prompt Injection: "Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection". Greshake et al. AISec 2023. [Paper]
• AgentHarm†: "AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents". Andriushchenko et al. arXiv 2024. [Paper]
• TrustAgent: "TrustAgent: Towards Safe and Trustworthy LLM-Based Agents". Hua et al. EMNLP 2024. [Paper]
• ToolEmu†: "Identifying the Risks of LM Agents with an LM-Emulated Sandbox". Ruan et al. arXiv 2023. [Paper]
• Ignore Previous Prompt: "Ignore Previous Prompt: Attack Techniques For Language Models". Perez & Ribeiro. NeurIPS ML Safety Workshop 2022. [Paper]

Evaluation & Benchmarking

Key numbers: HAL unified 21,730 rollouts, compressing weeks of evaluation to hours; OSWorld reports 28% false negative rate in automated evaluation; METR finds benchmark-passing PRs have 24.2pp lower human merge rate, widening at 9.6pp/year.

• AgencyBench†: "AgencyBench: Benchmarking the Frontiers of Autonomous Agents in 1M-Token Real-World Contexts". Li et al. arXiv 2026. [Paper]
• AEGIS†: "AEGIS: No Tool Call Left Unchecked -- A Pre-Execution Firewall and Audit Layer for AI Agents". Yuan et al. arXiv 2026. [Paper]
• Hell or High Water†: "Hell or High Water: Evaluating Agentic Recovery from External Failures". Wang et al. COLM 2025. [Paper]
• SearchLLM†: "Aligning Large Language Models with Searcher Preferences". Wu et al. arXiv 2026. [Paper]
• Meta-Harness†: "Meta-Harness: End-to-End Optimization of Model Harnesses". Lee et al. arXiv 2026. [Paper]
• TheAgentCompany†: "TheAgentCompany: Benchmarking LLM Agents on Consequential Real-World Tasks". Xu et al. arXiv 2024. [Paper]
• BrowserGym†: "The BrowserGym Ecosystem for Web Agent Research". Le Sellier De Chezelles et al. arXiv 2024. [Paper]
• WorkArena†: "WorkArena: How Capable are Web Agents at Solving Common Knowledge Work Tasks?". Drouin et al. arXiv 2024. [Paper]
• R-Judge: "R-Judge: Benchmarking Safety Risk Awareness for LLM Agents". Yuan et al. EMNLP 2024. [Paper]
• R2E: "R2E: Turning any GitHub Repository into a Programming Agent Environment". Jain et al. ICML 2024. [Paper]
• Evaluation Survey: "Evaluation and Benchmarking of LLM Agents: A Survey". Mohammadi et al. KDD 2025. [Paper]
• PentestJudge†: "PentestJudge: Judging Agent Behavior Against Operational Requirements". Caldwell et al. arXiv 2025. [Paper]

Protocol Standardization

Key numbers: MCP (tool↔harness): 2–15ms latency; A2A (agent↔agent): 50–200ms; ACP (intent-level, IBM) — three protocols serve complementary roles.

• MCP: "Model Context Protocol". Anthropic. Technical Report 2024. [Spec]
• A2A: "Agent-to-Agent Protocol". Google. Technical Report 2025. [Spec]
• Protocol Comparison†: "A Survey of Agent Interoperability Protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP)". Ehtesham et al. arXiv 2025. [Paper]
• Gorilla: "Gorilla: Large Language Model Connected with Massive APIs". Patil et al. NeurIPS 2023. [Paper] [Code]

Runtime Context Management

Key numbers: SkillsBench — curated skill injection yields +16.2pp improvement; "Lost in the Middle" effect documented; long-context models shift the problem from retention to salience.

• SkillsBench†: "SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks". Li et al. arXiv 2026. [Paper]
• ReadAgent: "A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts". Lee et al. ICML 2024. [Paper]
• MemoryOS: "Memory OS of AI Agent". Kang et al. arXiv 2025. [Paper]
• CoALA: "Cognitive Architectures for Language Agents". Sumers et al. TMLR 2024. [Paper]
• SkillFortify†: "Formal Analysis and Supply Chain Security for Agentic AI Skills". Bhardwaj. arXiv 2026. [Paper]
• Lost in the Middle: "Lost in the Middle: How Language Models Use Long Contexts". Liu et al. TACL 2024. [Paper]
• Context Engineering Survey†: "Context Engineering: A Survey of 1,400 Papers on Effective Context Management for LLM Agents". Mei et al. arXiv 2025. [Paper]

Tool Use & Registry

Key numbers: Vercel found removing 80% of tools helped more than any model upgrade; Schema First (Sigdel & Baral, 2026) — a controlled pilot showing that schema-based tool contracts reduce interface misuse but not semantic misuse, with end-task success at zero across all conditions, suggesting interface design alone is insufficient for tool reliability; CodeAct outperforms on 17/17 Mint benchmarks with −20% turns.

• CodeAct: "Executable Code Actions Elicit Better LLM Agents". Wang et al. ICML 2024. [Paper] [Code]
• Schema First†: "Schema First Tool APIs for LLM Agents: A Controlled Study of Tool Misuse, Recovery, and Budgeted Performance". Sigdel & Baral. arXiv 2026. [Paper]
• ToolLLM: "ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs". Qin et al. ICLR 2024. [Paper] [Code]
• ToolSandbox†: "ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities". Lu et al. arXiv 2024. [Paper]
• AutoTool†: "AutoTool: Efficient Tool Selection for Large Language Model Agents". Jia & Li. AAAI 2026. [Paper]
• Tool Learning Survey: "Tool Learning with Large Language Models: A Survey". Qu et al. Frontiers of Computer Science 2024. [Paper]
• GoEX†: "GoEX: Perspectives and Designs Towards a Runtime for Autonomous LLM Applications". Patil et al. arXiv 2024. [Paper]
• AgentTuning: "AgentTuning: Enabling Generalized Agent Abilities for LLMs". Zeng et al. ACL 2024. [Paper]

Memory Architecture

Key numbers: Mem0 achieves 90% token reduction vs full-context; Zep temporal knowledge: +18.5% QA accuracy; Agent Workflow Memory: +14.9% on Mind2Web. Six architectural patterns: flat buffer → hierarchical → episodic → semantic → procedural → graph.

• Agent Workflow Memory (AWM)†: "Agent Workflow Memory". Wang et al. arXiv 2024. [Paper]
• Mem0†: "Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory". Khant et al. arXiv 2025. [Paper]
• A-MEM†: "A-MEM: Agentic Memory for LLM Agents". Xu et al. NeurIPS 2025. [Paper]
• MemAct†: "Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks". Zhang et al. arXiv 2025. [Paper]
• Memory Survey†: "Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers". Du. arXiv 2026. [Paper]
• MemoryBank: "MemoryBank: Enhancing Large Language Models with Long-Term Memory". Zhong et al. AAAI 2024. [Paper]
• LoCoMo†: "Evaluating Very Long-Term Conversational Memory of LLM Agents". Maharana et al. arXiv 2024. [Paper]
• Memory Mechanisms Survey†: "A Survey on the Memory Mechanism of Large Language Model Based Agents". Zhang et al. arXiv 2024. [Paper]
• Evo-Memory†: "Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory". Wei et al. arXiv 2025. [Paper]

Planning & Reasoning

Key numbers: SWE-agent ACI study shows interface design outweighs model capability as the primary performance determinant. LATS integrates MCTS with language model feedback for state-space search. Plan-on-Graph enables adaptive self-correcting planning on knowledge graphs through guidance, memory, and reflection mechanisms.

• Tree of Thoughts: "Tree of Thoughts: Deliberate Problem Solving with Large Language Models". Yao et al. NeurIPS 2023. [Paper] [Code]
• LATS: "Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models". Zhou et al. arXiv 2023. [Paper] [Code]
• Plan-on-Graph: "Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs". Chen et al. NeurIPS 2024. [Paper]
• AFlow†: "AFlow: Automating Agentic Workflow Generation". Zhang et al. arXiv 2024. [Paper]
• Agent Q†: "Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents". Putta et al. arXiv 2024. [Paper]
• OPENDEV†: "Building Effective AI Coding Agents for the Terminal: Scaffolding, Harness, Context Engineering, and Lessons Learned". Bui. arXiv 2026. [Paper]
• AOrchestra†: "AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration". Ruan et al. arXiv 2026. [Paper]
• RAP: "Reasoning with Language Model is Planning with World Model". Hao et al. EMNLP 2023. [Paper]
• Inner Monologue: "Inner Monologue: Embodied Reasoning Through Planning with Language Models". Huang et al. CoRL 2022. [Paper]
• Agent-Oriented Planning: "Agent-Oriented Planning in Multi-Agent Systems". Li et al. ICLR 2025. [Paper]
• ExACT†: "ExACT: Teaching AI Agents to Explore with Reflective-MCTS and Exploratory Learning". Yu et al. arXiv 2024. [Paper]

Multi-Agent Coordination

Key numbers: AgencyBench — agents achieve 48.4% success on native SDK harness vs substantially lower on independent harnesses, demonstrating tight harness-agent coupling. Byzantine fault tolerance remains an open problem for adversarial multi-agent settings.

• SAGA†: "SAGA: A Security Architecture for Governing AI Agentic Systems". Syros et al. NDSS 2026. [Paper]
• MAS-FIRE†: "MAS-FIRE: Fault Injection and Reliability Evaluation for LLM-Based Multi-Agent Systems". Jia et al. arXiv 2026. [Paper]
• Byzantine fault tolerance†: "Rethinking the Reliability of Multi-agent System: A Perspective from Byzantine Fault Tolerance". Zheng et al. arXiv 2025. [Paper]
• Multi-agent baseline study†: "Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline". Xu et al. arXiv 2026. [Paper]
• AgentVerse†: "AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors". Chen et al. arXiv 2023. [Paper]
• Mixture-of-Agents†: "Mixture-of-Agents Enhances Large Language Model Capabilities". Wang et al. arXiv 2024. [Paper]
• Multi-Agent Survey: "Large Language Model Based Multi-Agents: A Survey of Progress and Challenges". Guo et al. IJCAI 2024. [Paper]
• Concordia†: "Generative Agent-Based Modeling with Actions Grounded in Physical, Social, or Digital Space Using Concordia". Vezhnevets et al. arXiv 2023. [Paper]

Compute Economics

Key numbers: OpenRouter reports 13T tokens/week (Feb 2026), doubling every 4 weeks; AgencyBench measures 1M tokens/task average; 1000× agent compute growth projected by 2027; AIOS achieves 2.1× throughput speedup via proper agent scheduling.

• Repo2Run†: "Repo2Run: Automated Building Executable Environment for Code Repository at Scale". Hu et al. arXiv 2025. [Paper]
• Policy-First†: "Guardrails as Infrastructure: Policy-First Control for Tool-Orchestrated Workflows". Sigdel & Baral. arXiv 2026. [Paper]

---

Emerging Topics

• Self-Evolving Agents Survey†: "A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence". Gao et al. TMLR 2026. [Paper]
• Self-RAG: "Self-RAG: Learning to Retrieve, Generate, and Critique Through Self-Reflection". Asai et al. ICLR 2024. [Paper]
• Constitutional AI: "Constitutional AI: Harmlessness from AI Feedback". Bai et al. arXiv 2022. [Paper]
• AppAgent†: "AppAgent: Multimodal Agents as Smartphone Users". Zhang et al. arXiv 2023. [Paper]

Related Surveys

• LLM Agents Survey: "A Survey on Large Language Model Based Autonomous Agents". Wang et al. arXiv 2023. [Paper]
• Rise of LLM Agents: "The Rise and Potential of Large Language Model Based Agents: A Survey". Xi et al. arXiv 2023. [Paper]
• LLM Survey: "A Survey of Large Language Models". Zhao et al. arXiv 2023. [Paper]
• AI Agent Systems†: "AI Agent Systems: Architectures, Applications, and Evaluation". Xu. arXiv 2025. [Paper]

Practitioner Reports & Industry Insights

Production deployment experiences from Stripe, OpenAI, Cursor, METR, and other frontier practitioners.

• Stripe Minions: "Minions: Stripe's one-shot, end-to-end coding agents". Gray. Stripe Dev Blog, Feb 2026. [Blog]
• Harness Engineering (OpenAI): "Harness engineering: leveraging Codex in an agent-first world". Lopopolo. OpenAI Blog, Feb 2026. [Blog]
• Self-Driving Codebases: "Towards self-driving codebases". Lin. Cursor Blog, Feb 2026. [Blog]
• METR SWE-bench Analysis: "Many SWE-bench-Passing PRs Would Not Be Merged into Main". Whitfill et al. METR Notes, Mar 2026. [Report]

---

Future Directions

Eight open research directions identified in the survey (no curated paper list — these are forward-looking):

• Formal Harness Specification Language — DSL for specifying and verifying H=(E,T,C,S,L,V) components
• Cross-Harness Benchmark Suite — portability testing across incompatible harness ecosystems
• Security Taxonomy & Threat Model — extension of OWASP Top 10 to agent harness attack surfaces
• Protocol Interoperability (MCP/A2A) — bridging tool-level and agent-level protocols
• Long-Horizon Evaluation Methodology — metrics that don't degrade under multi-session, multi-day tasks
• Harness-Aware Fine-Tuning — training models that are aware of their execution environment
• Memory Interface Standardization — portable memory APIs across flat, episodic, and graph stores
• Harness Transparency Specification — auditability and explainability for runtime decisions

---

Citation

See BibTeX at the top of this README.

---

Update Log

Version	Date	Changes
v1	2026-04-03	Initial preprint
v2	2026-04-07	Repo updated

---

† denotes preprint, not yet peer-reviewed.
This survey is under active development; the full manuscript will be released soon.
Maintained by Qianyu Meng & Liyi Chen. PRs welcome for missing papers or updated links.