AI-Agents Group: Market Research on Agent Platforms and Frameworks (Group Overview and Cross-Platform Comparison)

1. Group Overview

1.1. Research Scope

This group conducts market research on the 21 most representative agent platforms / frameworks on the 2026 market, covering three types of vendor forms:

1. Native Harness from model vendors: Claude Agent SDK (Anthropic), OpenAI Agents SDK (OpenAI), Google ADK / Gemini CLI (Google), Microsoft Agent Framework (Microsoft), Mistral Agents API (Mistral), OpenAI Agents API (OpenAI, public beta 2026-09-10, Grade A official basis);
2. Frameworks and open-source platforms: LangGraph / LangChain / LangSmith, Dify, Coze / Kouzi (ByteDance), n8n, CrewAI, AutoGen and Semantic Kernel (Microsoft legacy line), AgentScope (Alibaba open-source foundation), OpenHands (All Hands AI);
3. Closed-source vertical products and cloud-vendor platforms: Devin (Cognition), Manus (Butterfly Effect), Tencent Yuanqi / Yuanbao, Alibaba Cloud Bailian, AWS Bedrock AgentCore (AWS), Replit Agent (Replit), HiAgent 3.0 (Volcano Engine), Huawei Cloud AgentArts (Huawei Cloud).

Each platform has one independent document (see Section 5, Document Navigation), uniformly following an eight-part structure: introduction, terminology, feature description, platform architecture, Harness design (layer-by-layer analysis of the six layers), real-world cases, summary, and references.

1.2. Analysis Framework and Rating Criteria

图 1-1｜Harness 六层能力模型：21 个智能体平台的统一评级框架

数据来源：基于本文分析绘制的示意图。

All platforms uniformly apply the Harness six-layer capability model defined in the project parameter card:

Rating criteria (four levels): Strong (native first-class citizen mechanism, configurable and extensible) / Medium-Strong (has a native mechanism but incomplete coverage, or a key capability is provided only in the commercial edition) / Medium (requires the user to assemble it themselves, or provides only basic primitives) / Weak (largely absent, needs to be filled in by external systems).

Methodology statement that must be emphasized: All six-layer ratings in this group are qualitative judgments based on publicly available materials, not unified-benchmark testing. There is no uniform performance / reliability benchmark across the 21 platforms (the only referable cross-platform data are Manus's GAIA results, CrewAI's independent completion-rate benchmark, and Devin's launch SWE-bench score, and these are not comparable to one another); the ratings reflect "the density and depth of publicly verifiable mechanisms", not "production-environment measured performance".

1.3. Lineage Positioning: Coordinates of the 21 Platforms in the Harness Lineage

According to the boundary definitions in the parameter card (Agent Framework is a subset of Harness, and Agent Platform is the productized packaging of Harness), the lineage coordinates of the 21 platforms are as follows:

• Beyond the summary positioning, several key events in 2026 further pin down each platform's actual position; see Section 1.4.

1.4. 2026 Key Events Timeline

The following events directly affected the competitive landscape of the 21 platforms in this group, and each document expands on its corresponding one:

---

2. Cross-Platform Comparison Matrix of the 21 Platforms

2.1. Basic Information and Positioning Matrix

2.2. Harness Six-Layer Maturity Matrix

Ratings are qualitative judgments based on publicly available materials, not unified-benchmark testing; "Weak" means largely absent and needing external fill-in, and "(Opaque)" means closed-source and unverifiable.

Three structural readings of the matrix:

1. Three all-six-layers-strong platforms: Microsoft Agent Framework, Alibaba Cloud Bailian, and AWS Bedrock AgentCore — the first two rely respectively on enterprise-governance depth and the "platform + open-source foundation" dual track, while AgentCore relies on the "six-layer-capability componentization" cloud-service form, and is the only all-strong implementation not bound to a single model;
2. No platform is all-strong on L5 and free and open-source: the evaluation-observability layer is either locked in commercial editions (LangSmith, AMP) or on cloud platforms (Bailian; OpenAI Evals already announced for shutdown), making L5 the most obvious public gap in the current market;
3. There are two kinds of "Strong" at L2: Claude Agent SDK / ADK are strong in "giving the agent a computer" (sandbox and permissions), while n8n is strong in "letting the agent reach everything" (connector density) — the same Strong rating has different implications.

2.3. License Comparison

2.4. Quick Reference: "Strongest Implementation" of Each Layer

---

3. Competitive Focal Points: Three Intrinsic Tensions

The product divergence among the 21 platforms essentially reflects choosing different positions along three tension axes. These three tensions are the underlying coordinate system for understanding all of this group's comparison conclusions.

Tension One: Flexibility ↔ Predictability

The higher an agent's autonomy, the more unpredictable its execution path; the stronger the determinism, the narrower the range of tasks it can handle. Where each platform lands:

• Prediction-oriented: Dify (officially recommends demoting the Agent to a workflow node), Bailian / CrewAI (a design of deterministic containers housing a probabilistic kernel), n8n (DAG semantics);
• Flexibility-oriented: Manus and Devin (end-to-end autonomous delivery), Coze Agent World (dynamic multi-agent routing), the Agent 2.0 planning engine (autonomously deciding tool calls);
• Middle ground with explicitly managed tension: LangGraph (interrupt + explicit Command pause points), Claude Agent SDK (Plan mode separating planning from execution).

Tension Two: Openness ↔ Governance

The more open the ecosystem (multi-model, MCP, community plugins, third-party agent integration), the greater the attack surface and governance difficulty. Where each platform lands:

• Openness-oriented: OpenAI Agents SDK (LiteLLM connecting 100+ providers), Bailian (150+ models + bidirectional MCP + open-source foundation), Coze (integrating external agents such as Claude Code / Codex CLI);
• Governance-oriented: Microsoft Agent Framework (the full Middleware / Filters / Entra / sandbox set), Claude Agent SDK (layered permissions + budget guardrails), Bailian (three-state permissions + shell static analysis);
• Closedness for certainty: Tencent Yuanqi (Tencent-ecosystem channel loop), Devin / Manus (closed-source product contracts).

Tension Three: Cost ↔ Depth

The deeper an agent goes (multi-agent, long tasks, heavy retrieval), the greater the token and compute consumption, and the worse the predictability of that consumption. Where each platform lands:

• Cost-friendly: n8n (free self-hosting + per-execution billing), open-source frameworks (MIT / Apache 2.0 self-hosting);
• Depth-expensive: Devin (ACU billing, high and unpredictable unit cost), Manus (credit consumption of 500—900 per complex task with no cap), enterprise-edition governance (Dify / CrewAI / Bailian's key capabilities sit in paid tiers);
• Mitigation approaches converge: tool-result compression (Bailian), model routing (mixing models inside Dify workflows), budget guardrails (Claude Agent SDK's max_budget_usd).

---

4. Selection Recommendations

4.1. Quick Selection Reference by Scenario

4.2. Selection by Organization Form

• Engineering teams building in-house: prefer Runtime-type SDKs (Claude / OpenAI / ADK / MAF), choosing based on model and cloud-ecosystem lock-in;
• Platform teams operating an AI platform: prefer Agent Platform (Dify / Coze / Bailian / Yuanqi), choosing based on privatization requirements, ecosystem channels, and compliance requirements;
• Regulated industries: prefer solutions strong on L6 (MAF, Bailian, Claude Agent SDK), and confirm that governance capabilities are not behind a paywall;
• Business teams without an engineering team: prefer no-code platforms (Coze, Yuanqi, Dify), and budget human resources for the L5 evaluation gap.

4.3. Combination Strategies

In production, a single platform covering everything is rare; the mature combination patterns that recur throughout this group's research are:

1. Framework + connectivity layer: LangGraph / CrewAI / Claude Agent SDK handle reasoning orchestration, while n8n handles triggering, integration, and write-back;
2. Platform + evaluation layer: Dify / Coze handle building and publishing, with LangSmith / Coze Loop / Langfuse filling in L5;
3. Open-source foundation + commercial cloud: verify with AgentScope open source, then scale up with Bailian on the cloud; self-host Coze Studio, and add governance with Volcano Engine enterprise solutions;
4. Vertical product + general platform: Devin takes on bounded engineering tasks, while the general platform takes on business agents.

4.4. Pre-Adoption Checklist

Regardless of which platform you choose, review each item below one by one before going live (the information-gap statements in each document provide the corresponding verification entry points):

1. License and business-model compatibility: does your usage touch any license red lines (Dify's multi-tenancy prohibition, n8n's fair-code threshold, Claude CLI commercial terms);
2. Whether governance capabilities sit behind the paywall: are SSO / RBAC / auditing / privatization available at your chosen tier (Dify, CrewAI, n8n, and Bailian all have paid tiers);
3. L5 evaluation plan: is the platform's native evaluation sufficient (Bailian / LangGraph / OpenAI), otherwise are the external options (LangSmith / Coze Loop / Langfuse) and their data-export capabilities feasible (Tencent Yuanqi's call-data exportability has not been publicly documented);
4. Long-task reliability: do you need persistent execution and resumable checkpoints (LangGraph / ADK / MAF / Bailian have it; n8n / Dify / Yuanqi lack it or have not disclosed it);
5. Cost predictability: does the billing unit (execution / token / credit / ACU) match your load profile, and do you need budget guardrails (Devin and Manus have the largest consumption fluctuation);
6. Model lock-in risk: the depth of support for non-default models (Claude Agent SDK supports only Claude; Bailian is centered on Qwen; Yuanqi is centered on Hunyuan);
7. Vendor-survival risk: funding and organizational changes (Dify's team reduction, Manus's independent operations), product shutdowns (OpenAI Agent Builder / Evals discontinued), and ecosystem shifts (Gemini CLI → Antigravity);
8. Version currency: versions and feature descriptions are based on the September 2026 retrieval point; re-verify against official documentation before adoption.

---

5. Document Navigation

Note: The six-layer ratings for rows 01—21 are back-filled from the main text of each volume's "Harness Design" section; when a volume revises its ratings, this table must be updated in sync.

---

6. Research Limitations and Global Information Gaps

This group's research has the following global limitations for cross-platform comparison; you should read this section before citing any conclusion:

1. Six-layer ratings are not unified-benchmark testing: all ratings are qualitative judgments based on publicly available materials. There is no unified comparable benchmark across the 21 platforms; the only referable cross-platform data are Manus's GAIA results (L1 86.5% / L2 70.1% / L3 57.7%), CrewAI's complex multi-step task completion rate of roughly mid-50% in independent 2026 benchmarks, and Devin's launch SWE-bench score of 13.86%, and these three are not comparable to one another.
2. Quantitative effectiveness data are almost entirely vendor-reported: of the 21 platforms, only Dify (Sichuan Rural Commercial Bank's anti-money-laundering assistant) and OpenAI Agents API (SafetyKit / Hypha / Cirridae / Nash early-customer data) have public cases found with engineering metrics; the vast majority of the rest are vendor-reported or media-transmitted and have not been independently verified by third parties.
3. Closed-source products cannot be verified: Devin's and Manus's L1 (context engineering), L4 (memory & state), and L5 (evaluation & observability) mechanisms are closed-source and unverifiable, so the related ratings can only be inferred from product behavior; each document explicitly marks this as "mechanism not disclosed".
4. Domestic platforms have low public technical documentation: Tencent Yuanqi's architecture, state management, persistence, and evaluation system have no official public documentation at all, so its ratings are constrained by "insufficient public information"; Coze's architecture information comes mainly from third-party reverse engineering and deployment testing; the official public availability of HiAgent 3.0's and Huawei Cloud AgentArts's architecture, memory, and evaluation details is similarly low, and the two documents' density of / [To be filled] is significantly above average.
5. Conflicting figures are each preserved: three Hunyuan version figures (2.0 / Hy3 / Hy4 preview) coexist, two claims about Yuanqi privatization, two claims about CrewAI's operating scale and adoption rate, multiple claims about Coze stars and plugin counts, and two claims about the Mistral Agents API launch date (2025-05-27 mainstream evidence vs. 2026-04/05 in a few third-party directories) all coexist; the documents do not forcibly merge them, so be careful when citing.
6. Versions are highly time-sensitive: all version numbers and feature availability are based on the 2026-09-11/12 retrieval point; leading projects iterate extremely fast (OpenAI Agents SDK about weekly, ADK about every two weeks), so cite the retrieval date when referencing. Known upcoming changes include: OpenAI Agent Builder and Evals will no longer be provided on the platform from 2026-11-30, Semantic Kernel maintenance mode lasts at least until 2027-04, and the Gemini CLI personal free tier migrated to Antigravity from 2026-06-18.
7. Product-survival and geopolitical risk: Manus went through Meta's acquisition being revoked by China's NDRC (2026-04) and resumed independent operations in 2026-08; Dify's R&D team dropped from 40 to 15 people in 2026. Long-term dependency assessment must include organizational-change factors.

---

7. References

1. 2026 Enterprise Agent Development Platform Panorama Review: Cross-Platform Comparison of Eight Mainstream Platforms — Juejin. https://juejin.cn/post/7654244323158016038
2. 2026 AI Agent Platform Full-Dimension Review: From "Lobster Farming" to Enterprise-Grade Deployment — CSDN. https://blog.csdn.net/weixin_56622231/article/details/159515126
3. Overview of Leading Domestic Enterprise Agent Companies: Orchestration, Governance, and Implementation by Six Companies Including Lenovo Group — China Economic Net. https://www.cet.com.cn/itpd/itxw/10523920.shtml
4. The Most Complete, Newest, and Most Detailed Panorama Guide to Tencent AI Products — CSDN. https://blog.csdn.net/weixin_49076592/article/details/163128796
5. Introducing AgentKit — OpenAI official blog. https://openai.com/blog/introducing-agentkit
6. Microsoft Agent Framework — Microsoft Learn official documentation. https://learn.microsoft.com/en-us/agent-framework/overview/
7. AgentScope — official website. https://agentscope.io
8. LangGraph: Agent Orchestration Framework — LangChain official website. https://www.langchain.com/langgraph
9. n8n official website. https://n8n.io/
10. Dify — official website. https://dify.ai/
11. Tencent Yuanqi — official website. http://yuanqi.tencent.com
12. Project Parameter Card v1.0 (Harness six-layer capability model and conceptual boundaries) — internal baseline document of this project.
13. R09-AI-Agents Group Retrieval Report — internal retrieval report of this project.