When to open:You want to understand how Agents are implemented from an AWS perspective, how the ANP protocol challenges MCP, and how independent developers ship products. This wiki is a consolidated compilation of 4 talks shared by AWS on the newtype platform in 2025.
Relationship with AI Agent Technical Architecture Evolution 2025 (In-depth Research Report) relationship:That document is an "industry landscape research report", while this wiki is the industry practical perspective from 4 specific talks.
I. Four Stages of Agent Evolution (Wang Qixiang, Solutions Architect at Amazon Web Services)
Core judgment: Agents are currently in the transition phase from L2 to L3.
| Stage | Description | Key Characteristics |
|---|---|---|
| L1 Single Agent w/ multiple tools | Instruction Executor | One LLM + tool stack |
| L2 Multi-Agent Collaboration | Participant with reasoning, self-reflection, planning, and correction capabilities | Multi-Agent collaboration |
| L3 Asynchronous Agents | Asynchronous parallel Agents | Task-driven rather than conversation-driven |
| L4 Autonomous Agents | Agents create Agents themselves, RL optimizes Agentic Workflow | Autonomous evolution |
Present (2025):Most production deployments are at L1-L2, L3 is seeing breakthroughs, and L4 remains in the research paper stage.
II. Four Elements of Agents (Wang Xiaoyan, Senior Strategic Advisor for Amazon Startup Ecosystem)
Agent = Large Language Model + (Observation + Thinking + Action + Memory), split into 4 components in engineering:
2.1 Planning
- Claude
<think>Tools:Enable LLMs to think explicitly before acting - Applicable Scenarios:
- Tool output analysis (carefully processing outputs from previous tool calls before acting)
- Scenarios with many instructions/constraints (detailed guidelines + compliance validation)
- Sequential decision-making scenarios (each action builds on previous ones, with high cost of error)
2.2 Memory
- Short-term context + long-term persistence (corresponding to the 3 layers of MemoryOS)
2.3 Tool Use & Action
Three forms coexist:
- MCP(Model Context Protocol) — standardized tool integration
- Computer Use(Multimodal end-to-end fine-tuned Model/Agent) — GUI automation
- Multi-Agent collaboration — cross-Agent protocol (ANP discussed below)
2.4 Mainstream Implementation Frameworks
- OWL / LangManus / OpenManus / OpenHands / Manus (open source and commercial options coexist)
III. Three Tiers of Agent Capability (Wang Xiaoyan)
Counterintuitive judgment: Not all Agents require "end-to-end fine-tuning"; there is a 3-tier capability ladder:
| Tier | Description | When to Use |
|---|---|---|
| Tier One | Build vertical applications using general Agent capabilities (General Agent Roadmap) | Most scenarios |
| Tier Two | Fully unlock model capabilities (hard structure → soft structure) | Moderate complexity |
| Tier Three | End-to-end fine-tuned Agent(Reinforcement Fine-tuning, RFT, is the development paradigm for next-generation LLM Agents) | Complex + high-frequency + data-rich scenarios |
Key references:
- Anthropic "Building effective agents" — example: Claude Code is a truly AI Native development approach
- OpenAI "Defining agents" — example: DeepResearch is an agent built with end-to-end reinforcement fine-tuning on the o3 model
3.1 The Significance of RL
Lessons from DeepSeek:
- RL on strong base model works
- RL with sparse rewards works
- o1-level performance easier than expected
Algorithms:PPO / GRPO / DPO
Rubric Engineering — the next "prompt engineering":
- Focus on the art of reward function design
- It is necessary to both ensure evaluation criteria align with ultimate goals and guard against reward hacking
- Multi-layer validation: format compliance → logical coherence → result accuracy
- A good pipeline far outperforms algorithmic tricks
- Choosing between Offline RL and Online RL
IV. Agent Product Design (Wang Xiaoyan)
4.1 Novice Market (C-end, general public)
| Product | Data | Design Key Points |
|---|---|---|
| Perplexity(Best Answer Engine) | $100 million ARR, $18 billion valuation, 85% gross margin | Intent recognition + Web Search optimization + crawl and storage + first token latency + answer templates + followup |
| Lovable(Vibe Coding) | ARR grew from 0 to $17 million in 3 months, fastest growth in Europe | Natural language + image-to-code generation, direct panel edits via Visual Edits |
4.2 Professional Market (B-end, vertical)
| Product | Data | Design Key Points |
|---|---|---|
| Cursor(Best AI coding assistant) | 2x growth in 4 months, ARR > $150 million, $10 billion valuation | Cursor tab model (in-flow Next Action Prediction) + self-developed Cursor Small + Cursor Composer multi-file editing |
| Gamma(AI PPT) | 28 employees → tens of millions ARR + nearly 50 million users + already profitable | "Write like a doc, Present like a deck", automatic interface conversion, modular components + Block-based editing |
4.3 Agent Pricing Strategy
Service-as-a-Software(SaaS 2.0):
- User tiering: free traffic acquisition → Plus membership (pay for results) → Pro membership (explore model boundaries)
- Let your product benefit from improvements in model capabilities, rather than being overwhelmed by overflowing model capabilities
- If API prices drop,do not lower membership fees; add features instead
4.4 Three Curves
"Keep three curves in your mind!"
- Model capability growth curve
- Model cost decline curve
- AI Agent Moore's Law(combining the two)
V. Agent Brain: Reasoning Model Selection (Wang Qixiang)
5.1 DeepSeek R1 Family
| Model | Description |
|---|---|
| DeepSeek V3 | Pretrain Base |
| DeepSeek R1-Zero | Pure RL training transition model,aha moment emergence |
| DeepSeek R1 | V3 + intermediate high-quality reasoning + small amount of SFT optimization alignment |
| Distill Series | Qwen 1.5B/7B/14B/32B, Llama 8B/70B —— Distillation is more effective than direct RL training on small models |
Key Insight:The base model's own capabilities are very important (distillation is more efficient than training).
5.2 R1 Advantages vs Limitations
| Advantages | Limitations |
|---|---|
| Strong reasoning ability (excellent at STEM) | function calling / multi-turn / role-playing / JSON output Gaps exist |
| Transparency(displays thinking process) | Language confusion(only optimized for Chinese and English) |
| Creative text generation, emotional value | Not suitable for simple tasks — overthinking far outweighs accuracy gains, wasting large numbers of tokens |
| Does not support multimodality, poor safety guardrails |
5.3 Claude 3.7 Sonnet (Hybrid Reasoning)
The industry's first generally available hybrid reasoning model:Seamless switching between Standard mode (fast) and Extended Thinking (detailed analysis).
- SWE-bench Verified Standard mode 70.3%
- Supports 128K Token output
- Adjustable reasoning tokens (allocate how much thinking budget)
Extended Thinking applicable scenarios:Complex analysis / advanced STEM / long context / complex instructions / structured creativity / detailed data generation / constrained optimization tasks.
Counterexample:Using Extended Thinking for simple tasks does not bring significant performance improvement, but instead wastes tokens.
5.4 Reasoning Model Prompt Practices
- Open-ended instructions > step-by-step instructions:Reasoning models perform better on open-ended instructions; solutions developed autonomously by the model often surpass human presets
- N-shot Prompting remains effective — Show rather than Tell
- Three tricks to improve consistency:
- Require Claude to verify via simple tests before completion
- Require Claude to analyze whether the previous step met expectations
- Run multiple test cases during Extended Thinking for programming tasks
5.5 Model as Product Trend
| Observation | Implication |
|---|---|
| Reasoning model + RL Scaling is still in its early stages | No all-around SOTA, continuous attention required |
| Adjustable thinking depth becomes standard | Hybrid reasoning will dynamically allocate computing power in the future |
| FM providers see diminishing marginal returns → push Claude Code / Agentic products | Selling models is less profitable than selling applications |
| Vertical industries with private data | Build small models using the new RL + distillation paradigm |
VI. Multi-Agent Framework Comparison (Wang Qixiang)
| Framework | Architecture | Memory | Tool | Modality | Learning Curve | Scalability | Deployment | Best Scenario |
|---|---|---|---|---|---|---|---|---|
| AWS Multi-Agent Orchestrator | Cloud-native | In-Memory / RDS / DynamoDB | AWS Bedrock/Lex/Lambda | Text | Low (Python/AWS) | High - enterprise-grade | Low (Serverless) | Large-scale internal/external enterprise applications |
| AutoGen | Conversational agents | Session-base | Custom Python | Text/Image | Moderate | Low | High | Personal complex problems, experimental use |
| LangGraph | Graph-based stateful workflows | Global State / LangMem | LangChain ecosystem | Text/Image | Steep | Moderate | Medium | Small and medium-scale enterprise internal use |
| Dify | Low-code workflow builder | External tools | Marketplace plugins | Text/Image/Any | Very low | High (commercial version) | Moderate | Small and medium-scale use by non-technical teams |
| MetaGPT | Role-based SOP-driven | Session/Global env | Custom Python | Text | Moderate | Low | High | Programming SOP / data analysis |
6.1 AWS Multi-Agent Orchestrator Features
- Flexible orchestration:Task routing / Supervisor orchestration / sequential workflows / hybrid multi-level
- Built-in Session memory (In-Memory / SQL / DynamoDB)
- Integrates Bedrock Knowledge Base retrieval
- Serverless deployment (AWS Lambda)
- Modular: Agent / Tool / Storage / Retriever abstractions
6.2 MetaGPT Core Philosophy
Code = SOP(Team)
- Foundational Components Layer:Environment / Memory / Roles / Actions / Tools
- Collaboration Layer: Knowledge Sharing, each Agent maintains a message queue, triggers observe → think → action → publish via
_watch(action)message subscription
VII. ANP Protocol: Email for Agents (Chang Gaowei, Founder of the ANP Open Source Community)
Core thesis: MCP is the USB-C for models,ANP is the email for agents. The two do not conflict; they operate at different layers.
7.1 Four Ways AI/Agents Interact with the Digital World
| Method | Advantages | Disadvantages |
|---|---|---|
| API | High efficiency | Lack of standards, few APIs |
| Computer Use | Most abundant supply | High cost, poor experience |
| Browser Use | (Browser automation) | Cannot access APPs |
| Protocol | AI-native connection method, standardized | High transformation cost |
7.2 Agentic Web Trends
3 trends:
- Agents will replace existing APPs and software
- Interconnection and intercommunication must be achieved between Agents
- Agents connect and communicate via protocols
2 challenges:
- Convenient identity verification for all agents
- All agents can connect with other agents
7.3 What is ANP
Goal: Become the HTTP of the Agentic Internet era.
Vision:Build an open, secure, and efficient collaboration network for billions of agents.
Features:
- The world's first communication protocol designed for agents
- Enables any two agents to interconnect and communicate
- Builds a data network easily accessible to AI
7.4 ANP Layered Architecture
┌────────────────────┐
│ 应用协议层 │
├────────────────────┤
│ 元协议层 │
├────────────────────┤
│ 身份与加密通信层 │
└────────────────────┘
7.5 Three Core Problems Solved by ANP
- Agent identity: Decentralized (non-blockchain) identity authentication based on W3C DID. Agents can use their own IDs to perform identity authentication, communication, and collaboration with agents on any platform/organization
- Agent Description: Based on the Semantic Web (JSON-LD / schema.org), enables agents to publish basic information, capabilities, and interfaces, linking them into a network that is easy for AI to read
- Agent Discovery: Similar to DNS discovery for websites, search engines can discover all agents across the entire network
7.6 W3C DID
- Decentralized Identifier(Decentralized Identifier)
- Became a W3C Recommendation in 2022
- A user-controlled, fully self-sovereign digital identity identifier
- Typical applications:Bluesky
- did:wba method:Fully leverages existing Web infrastructure, can support billions of users at scale
7.7 Differences between MCP and ANP (Key!)
| Dimension | MCP | ANP |
|---|---|---|
| Analogy | USB-C for models | Email for agents |
| Good at | Model ↔ Tools / Resources | Agent ↔ Agent communication and collaboration |
| Not good at | Agent communication (each Agent must register an account) | Tool-type connections |
| Identity Authentication | OAuth 2.1(Centralized, requires account registration) | W3C DID(Decentralized, interact directly using own ID) |
| Protocol Architecture | CS (Client-Server, client initiates connection actively) | P2P(Peer-to-Peer, any party can initiate) |
| Information Organization | RPC remote invocation of resources/tools list | Linked-Data(URL links, reusing Web infrastructure) |
| Design Philosophy | Model-Centric (the Internet is the context and tools for models) | Agent-Centric(Every agent has equal status, a decentralized collaboration network) |
Key Conclusion:MCP is used to build agents, and agents use ANP for collaboration. The two are complementary.
7.8 Problems with MCP for Agents (Taking Personal Assistant + Multi-Agent as an Example)
- The personal assistant must register an account on every agent
- Agents cannot actively connect to the personal assistant
→ ANP solves this with W3C DID + P2P
7.9 ANP Case Studies
Open-source Manus (owl/OpenManus) integrates ANP:
- Interacts with agents supporting the ANP protocol via the ANP SDK
- Has developed hotel and attraction agents + weather query agent
Community Status: 20+ developers (Alibaba, Ant Group, ByteDance, Huawei, Kuaishou, iFLYTEK).
7.10 ANP Philosophy
Connection is power. As long as a person can freely connect to information, connect to others, and connect to tools, they have the ability to change the world.
Let the Internet return from platforms to connections.
VIII. Independent Development in Practice (idoubi Ai Doubi, former Tencent Senior Engineer)
8.1 Personal Background
- Former Tencent senior engineer, WeChat backend developer
- Currently freelancing,all in on AI application overseas expansion
- Full-stack: Go / PHP / Python / NodeJS / K8S
- Has launched over 10 AI products
8.2 Timeline of Released Products
| Time | Product | One-liner |
|---|---|---|
| 2023-02 | GPTalk | AI conversation assistant |
| 2023-07 | Zhiliao Reading (Cicada Reading) | AI summary |
| 2023-11 | GPTs Works | GPTs navigation site |
| 2023-12 | AI Wallpaper | AI wallpapers |
| 2024-01 | AI Cover | Red envelope covers |
| 2024-02 | Sora FM | AI video |
| 2024-03 | ThinkAny | AI search engine |
| 2024-04 | Melodisco | AI music |
| 2024-05 | HeyBeauty | Virtual try-on |
| 2024-08 | Pagen | AI landing pages |
| 2024-09 | CroSearch | Cross-language search |
| 2024-10 | PodLM | AI podcasts |
| 2024-11 | MCP.so | MCP application marketplace |
| 2024-12 | ShipAny | AI SaaS development framework |
| 2025-02 | CopyWeb | AI webpage replication |
8.3 6 Core Principles for Independent Development
- In martial arts, speed is the only unbreakable truth — Launch in one hour / Implement over a weekend / Deliver in 7 days
- Speed is no match for refinement; take the long-term view — Continuous polishing of GPTalk + Zhiliao Reading
- Dream big, start small — General vs vertical / Main tracks: large models / AI search / Agent
- Generate your own revenue; to forge iron, one must be strong oneself — Capital only adds icing on the cake. Capital looks at 4 points: team / competitive landscape / business model / growth rate
- Traffic is king; continuously build influence — Three platforms: Twitter / Jike / Xiaohongshu
- The most important thing about making products is having fun
8.4 AI Application Development SOP (Tech Stack)
Dependencies for launching quickly in 1 hour:
| Category | Selection |
|---|---|
| Development Foundation | TypeScript / React / NextJS |
| Login and registration | next-auth / clerk |
| Data storage | supabase / neon |
| Multilingual support | i18n / next-intl |
| Payment | stripe / lemonsqueezy / creem |
| File storage | s3 / r2 |
| Project deployment | cloudflare / vercel |
| Domain name | godaddy / namecheap |
Implementation Plan:
- Full-stack frameworks: NextJS / Remix / Nuxt
- Open-source templates: OpenSaaS / saas-starter
- Commercial templates: ShipAny / ShipFast / MkSaaS
- UI components: Shadcn / MagicUI / HeroUI
8.5 ProductHunt Ranking SOP
The fastest way for product cold start:
- Submit the product one week in advance
- Strive to be Featured
- Push for ranking (paid user acquisition allowed)
8.6 Programmatic SEO SOP
The lowest-cost, fastest-acting growth method:
- Data collection + cleaning
- AI summary + structured content
- Server-side rendering + clear webpage structure
- Automatically build pages for long-tail keywords
- Regularly update sitemap
8.7 AI Wrapper Guide (Wrapping to Ride Trends)
Operation Path:
- Discover trends: HF Space / GitHub / Google Trends
- Snatch domain names: Godaddy / Namecheap
- Implement features: replicate / fal / openrouter
- Acquire traffic: SEO / Reddit / Twitter
- Monetize traffic: subscriptions / ads
8.8 5 All-in Directions Recommended by idoubi
| Direction | Sub-direction |
|---|---|
| AI Coding | Coding Editor / Copilot / Coding Agent / Vertical Coding Agent |
| Agent | General / Vertical / SaaS to Agent |
| Agent Infra | Tools / Planning / Memory / Boilerplate / VM/Container / Auth / Payment |
| MCP | MCP Servers / Marketplace (mcp.so) / Atomic capability platform (mcprouter) / Consumer terminals |
| (Implicit) Overseas AI applications | NextJS full-stack + PH + programmatic SEO + Twitter traffic |
8.9 "This Is MCP" Written by idoubi
- MCP principle popular science
- MCP protocol analysis
- MCP server development
- MCP client development
- MCP use cases
- MCP ecosystem
IX. Combined Insights from the 4-Talk Compilation
9.1 Protocol Ecosystem Panorama
┌─ MCP ────────────────────┐
│ 模型 ↔ 工具/资源 │ CS架构 / OAuth2.1 / 中心化
├─ ANP ────────────────────┤
│ 智能体 ↔ 智能体 │ P2P架构 / W3C DID / 去中心化
├─ A2A(LFAIR)─────────────┤
│ 跨厂商 Agent 协作 │ (Linux Foundation)
└─ AGENTS.md / llms.txt ───┘
网站给 Agent 的"使用说明" (Karpathy 提议,见 [[newtype-karpathy-software-3]])
9.2 Framework Selection (Synthesized from 4 Talks)
| If you are | Recommended framework / protocol |
|---|---|
| Enterprise AWS user | AWS Multi-Agent Orchestrator + Bedrock |
| Personal complex Agent | AutoGen + custom |
| Non-technical team | Dify(low-code) |
| Programming SOP / data pipeline | MetaGPT |
| Graph state workflow | LangGraph |
| Want to do tool integration | MCP |
| Want to do cross-agent collaboration | ANP |
| Independent overseas expansion | NextJS + ShipAny + PH + programmatic SEO + Twitter |
9.3 Iron Rules for Agent Implementation (Consensus across 4 Talks)
- Capability tier matching scenarios: Not all Agents need end-to-end fine-tuning; general Agent + Prompt already covers most scenarios
- Differentiated model selection: Don't use Reasoning Models for simple tasks (overthinking wastes tokens)
- Three curves: Rising model capability ↑ + falling cost ↓ = Agent's Moore's Law; product design should benefit from this
- Service-as-a-Software pricing: Users pay for results; API price cuts do not lead to membership fee cuts
- Protocol layer division of labor: MCP handles tools, ANP handles agent collaboration
- Independent development prioritizes speed: Launch in 1 hour → deliver in one week; capital is not a necessity
9.4 Connections to Existing Wiki
- AI Agent Technology Architecture Evolution 2025 (In-depth Research Report) — Research report-level overview (this wiki takes an industry practical perspective)
- newtype · MCP Protocol and Ecosystem in Practice — newtype's judgment on MCP (this wiki supplements the ANP perspective)
- Karpathy · Software in the era of AI(Software 3.0) — Karpathy "Rewriting Infrastructure for Agents" (this wiki supplements ANP as an implementation example)
- Wang Kai Overseas Promotion SOP and Overseas Multi-Account Matrix in Practice — Wang Kai's overseas promotion (idoubi's SOP is a supplement from an independent developer's perspective)
History
- 2026-05-17: Phase 4 ingest created from 4 AWS Agent PDFs (01-04). Fully covers 4 stages of Agent + 4 elements + 3 capability tiers + Reasoning Model selection + Multi-Agent framework comparison + ANP protocol (benchmarked against MCP) + idoubi's independent development SOP.