Lesson 8: Agent Economy & Future — Full Spoken Transcript (Cantonese)

Original podcast: Cantonese dialogue between two speakers (M = 主持 host 子謙, F = 嘉賓 expert 曉晴). Total spoken duration: ~14 min. The audio above plays the full conversation. The text below is the complete transcript of what was actually said — preserved verbatim, not summarised, not translated.

Course: Eight-lesson course on AI agents · Lesson 8 of 8 · learnagent.lmmlab.com Topic (EN): marketplace · A2A protocol · payment rails · regulation · AGI timeline · multimodal agent. Topic (粵): marketplace、A2A、payment、regulation、AGI。 Speakers: 主持 M (host 子謙) and 嘉賓 F (expert 曉晴) · 27 spoken turns · ~14 min audio.

How to read this transcript

Each spoken turn from the podcast is shown as a separate paragraph, with the speaker label rendered in bold and an approximate timestamp (MM:SS) indicating where in the audio that turn occurs. Long turns are broken at natural sentence boundaries (。！？) and, where a sentence is still long, at clause boundaries (，；、). No English translation is provided — this is the exact spoken Cantonese audio transcript as it was synthesised for the lesson podcast.

Use browser Ctrl+F / ⌘+F to locate any phrase. On mobile (640 px viewport), the transcript scrolls as a single column.

Note: Sentence-level line breaks are for readability — the Cantonese text itself is preserved verbatim from the original podcast script (script_raw.json). No words have been altered, translated, summarised, or paraphrased.

Index of sections in this lesson

• 1. Opening & Agent Marketplace (粵: 開場同 Agent Marketplace) — turn 01 onwards
• 2. A2A Protocol & Agent Payment (粵: A2A Protocol 同 Agent Payment) — turn 05 onwards
• 3. Regulation & Economics (粵: Regulation 同 Economics) — turn 11 onwards
• 4. Future Trends & Skill Path (粵: Future Trends 同 Skill Path) — turn 17 onwards
• 5. Course Wrap-up (粵: 課程總結) — turn 23 onwards

Section 1/5 — Opening & Agent Marketplace

開場同 Agent Marketplace

Section overview: covers turns 01–04 (4 spoken segments).

Topic terms (extracted from spoken text): agent-to-agent, subscription, Marketplace, collaborate, reputation, autonomous, Agentforce, long-term, negotiate, landscape

Latin/English code-terms in this section (verbatim from speech): agent-to-agent, subscription, Marketplace

Section character total: 1,037 characters across 4 spoken turns.

Section duration estimate: ~2:04 of 14:00 total.

Turns in this section: 01, 02, 03, 04.

First spoken sentence of this section (turn 01, verbatim): 各位同學早晨, 我係子謙。

Average characters per turn (this section): ~259 chars.

Cumulative characters through this section: 1,037 of 9,528 total.

[01 | 00:00] 主持 M (host 子謙):

Turn 1 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 00:00 · section 1 (Opening & Agent Marketplace)

speaker=M · chars=201 · ts=00:00 · terms=agent-to-agent, Marketplace, reputation, autonomous, long-term · sentences=3 · clauses=10

Verbatim phrases in this turn: t Economy 同 Future, 包括 Agent Marketplace、 · A2A Protocol、 · agent-to-agent payment、

各位同學早晨, 我係子謙。歡迎收聽第八課, AI Agent 課程嘅最後一課。

First clause (verbatim): 各位同學早晨,

Last clause (verbatim): 同埋業界對 autonomous agent 嘅 long-term 預期。

各位同學早晨,

我係子謙。

歡迎收聽第八課,

AI Agent 課程嘅最後一課。

今日嘅主題係 Agent Economy 同 Future,

包括 Agent Marketplace、

A2A Protocol、

agent-to-agent payment、

agent identity 同 reputation system,

同埋業界對 autonomous agent 嘅 long-term 預期。

[02 | 00:31] 嘉賓 F (expert 曉晴):

Turn 2 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 00:31 · section 1 (Opening & Agent Marketplace)

speaker=F · chars=149 · ts=00:31 · terms=collaborate, autonomous, negotiate, landscape, discover · sentences=2 · clauses=9

Verbatim phrases in this turn: 佢哋會唔會形成自己嘅 economy, agent 之間點樣 discover、 · negotiate、 · ate、collaborate 同 transact, 呢個 landscape 點樣 evolve。

my, agent 之間點樣 discover、negotiate、collaborate 同 transact, 呢個 landscape 點樣 evolve。

First clause (verbatim): 大家好,

Last clause (verbatim): 呢個 landscape 點樣 evolve。

大家好,

我係曉晴。

今日嘅問題係,

當 agent 越來越 capable 同 autonomous,

佢哋會唔會形成自己嘅 economy,

agent 之間點樣 discover、

negotiate、

collaborate 同 transact,

呢個 landscape 點樣 evolve。

[03 | 01:02] 主持 M (host 子謙):

Turn 3 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 01:02 · section 1 (Opening & Agent Marketplace)

speaker=M · chars=319 · ts=01:02 · terms=subscription, Marketplace, developer, subscribe, platform · sentences=4 · clauses=8

Verbatim phrases in this turn: 首先講 Agent Marketplace 嘅 concept。 · t, 其他 developer 或者 user 可以 discover 同 use 呢啲 agent。 · mobile app store, 但 unit 唔係 mobile app 而係 AI agent。

首先講 Agent Marketplace 嘅 concept。

First clause (verbatim): 首先講 Agent Marketplace 嘅 concept。

Last clause (verbatim): billing 按 call 或者 subscription。

首先講 Agent Marketplace 嘅 concept。

Agent Marketplace 係 platform 讓 agent 開發者 publish 自己嘅 agent,

其他 developer 或者 user 可以 discover 同 use 呢啲 agent。

類似 mobile app store,

但 unit 唔係 mobile app 而係 AI agent。

例如一個 research agent 可以喺 marketplace 上面 listing,

其他 agent 可以 subscribe 同 use 佢做 research task,

billing 按 call 或者 subscription。

[04 | 01:33] 嘉賓 F (expert 曉晴):

Turn 4 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 01:33 · section 1 (Opening & Agent Marketplace)

speaker=F · chars=368 · ts=01:33 · terms=Marketplace, Agentforce, Anthropic, community, developer · sentences=7 · clauses=12

Verbatim phrases in this turn: 現有嘅 Agent Marketplace 例子。 · registry, developer 可以 browse 同 install MCP server。 · OpenAI 嘅 GPT Store, 雖然 2024 年開但 usage 唔算高。

現有嘅 Agent Marketplace 例子。

First clause (verbatim): 現有嘅 Agent Marketplace 例子。

Last clause (verbatim): 雖然唔專係 agent 但 host 好多 agent demo。

現有嘅 Agent Marketplace 例子。

Anthropic 嘅 MCP server marketplace,

包括 Smithery 同其他 community registry,

developer 可以 browse 同 install MCP server。

OpenAI 嘅 GPT Store,

雖然 2024 年開但 usage 唔算高。

Microsoft 嘅 Agent Catalog 喺 Microsoft Foundry 入面。

Google 嘅 Agent Garden 同 ADK 相關。

SalesHub 嘅 Agentforce,

客戶 service agent marketplace。

Hugging Face 嘅 Spaces,

雖然唔專係 agent 但 host 好多 agent demo。

End-of-section recap (last spoken sentence of Opening & Agent Marketplace): 現有嘅 Agent Marketplace 例子。

Section 2/5 — A2A Protocol & Agent Payment

A2A Protocol 同 Agent Payment

Section overview: covers turns 05–10 (6 spoken segments).

Topic terms (extracted from spoken text): agent-framework-a2a, Agent-of-agents, cross-framework, Auto-discovery, Agent-to-Agent, high-frequency, agent-friendly, cryptographic, decentralized, communication

Latin/English code-terms in this section (verbatim from speech): agent-framework-a2a, Agent-of-agents, cross-framework

Section character total: 2,241 characters across 6 spoken turns.

Section duration estimate: ~3:06 of 14:00 total.

Turns in this section: 05, 06, 07, 08, 09, 10.

First spoken sentence of this section (turn 05, verbatim): Agent discovery 嘅 mechanism。

Average characters per turn (this section): ~373 chars.

Cumulative characters through this section: 3,278 of 9,528 total.

[05 | 02:04] 主持 M (host 子謙):

Turn 5 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 02:04 · section 2 (A2A Protocol & Agent Payment)

speaker=M · chars=394 · ts=02:04 · terms=Agent-of-agents, Auto-discovery, description, capability, similarity · sentences=5 · clauses=14

Verbatim phrases in this turn: y, 即係 developer 將 agent metadata 例如 name、 · description、 · capability、

Agent discovery 嘅 mechanism。

First clause (verbatim): Agent discovery 嘅 mechanism。

Last clause (verbatim): SLA 保證。

Agent discovery 嘅 mechanism。

Static registry,

即係 developer 將 agent metadata 例如 name、

description、

capability、

pricing 上載到 registry,

其他 developer 查詢 registry 找適合嘅 agent。

Semantic search,

即係用 embedding similarity 配對 user query 同 agent capability。

Agent-of-agents,

即係 meta-agent 自己 discover 同 invoke 適合嘅 sub-agent 去 satisfy user 嘅 task。

Auto-discovery 嘅 challenge 包括 capability 描述嘅準確性、

pricing 透明度、

SLA 保證。

[06 | 02:35] 嘉賓 F (expert 曉晴):

Turn 6 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 02:35 · section 2 (A2A Protocol & Agent Payment)

speaker=F · chars=350 · ts=02:35 · terms=cryptographic, decentralized, application, reputation, historical · sentences=4 · clauses=13

Verbatim phrases in this turn: t 嘅 identity 唔同 human, 因為 agent 可以 clone、 · record, 即係 historical task success rate、 · user rating、

Agent identity 同 reputation system。

First clause (verbatim): Agent identity 同 reputation system。

Last clause (verbatim): decentralized reputation system 唔可以單方面篡改。

Agent identity 同 reputation system。

Agent 嘅 identity 唔同 human,

因為 agent 可以 clone、

fork、

modify,

所以 identity 應該 base on cryptographic key 而唔係 name。

Agent 嘅 reputation 需要 base on track record,

即係 historical task success rate、

user rating、

peer agent review。

區塊鏈技術嘅 application 喺呢度,

例如 immutable record 嘅 agent action,

decentralized reputation system 唔可以單方面篡改。

[07 | 03:06] 主持 M (host 子謙):

Turn 7 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 03:06 · section 2 (A2A Protocol & Agent Payment)

speaker=M · chars=381 · ts=03:06 · terms=agent-framework-a2a, cross-framework, Agent-to-Agent, communication, cross-runtime · sentences=5 · clauses=11

Verbatim phrases in this turn: A2A 嘅 design 目標係 cross-vendor、 · cross-framework、 · A2A Protocol, 即 Agent-to-Agent Protocol。

A2A Protocol, 即 Agent-to-Agent Protocol。

First clause (verbatim): A2A Protocol,

Last clause (verbatim): 透過 A2A 嘅 standard interface。

A2A Protocol,

即 Agent-to-Agent Protocol。

A2A 係 Google 同合作伙伴 2025 年提出嘅 protocol,

標準化 agent-to-agent communication。

Microsoft 嘅 Microsoft Agent Framework 透過 agent-framework-a2a adapter package 支援 A2A beta,

1.0 預計短期內推出。

A2A 嘅 design 目標係 cross-vendor、

cross-framework、

cross-runtime agent 可以 discover 同 collaborate。

例如 LangGraph 嘅 agent 可以 call CrewAI 嘅 agent,

透過 A2A 嘅 standard interface。

[08 | 03:37] 嘉賓 F (expert 曉晴):

Turn 8 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 03:37 · section 2 (A2A Protocol & Agent Payment)

speaker=F · chars=460 · ts=03:37 · terms=long-running, philosophy, capability, delegation, structured · sentences=6 · clauses=11

Verbatim phrases in this turn: nt 之間 exchange structured output 例如 JSON、 · A2A 嘅 technical detail。 · implement JSON-RPC 2.0, 類似 MCP 嘅 design philosophy。

A2A 嘅 technical detail。

First clause (verbatim): A2A 嘅 technical detail。

Last clause (verbatim): file。

A2A 嘅 technical detail。

A2A 喺 HTTP 上面 implement JSON-RPC 2.0,

類似 MCP 嘅 design philosophy。

A2A 嘅 capability 包括 agent discovery 透過 agent card,

即係 metadata document 描述 agent 嘅 capability 同 contact endpoint。

Task delegation 即係 parent agent 可以 delegate task 俾 child agent,

包括 task spec 同 expected output。

Status update 即係 child agent 喺 long-running task 上面 push progress update 俾 parent agent。

Artifact exchange 即係 agent 之間 exchange structured output 例如 JSON、

image、

file。

[09 | 04:08] 主持 M (host 子謙):

Turn 9 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 04:08 · section 2 (A2A Protocol & Agent Payment)

speaker=M · chars=341 · ts=04:08 · terms=Agent-to-agent, high-frequency, pay-per-call, Subscription, Micropayment · sentences=5 · clauses=12

Verbatim phrases in this turn: Agent-to-agent payment 嘅 model。 · 係每個 agent invocation charge fixed fee, 例如 0.001 美金。 · ription, 即係 monthly fee 包括 unlimited call 直至 quota。

Agent-to-agent payment 嘅 model。

First clause (verbatim): Agent-to-agent payment 嘅 model。

Last clause (verbatim): 適合 high-frequency agent interaction。

Agent-to-agent payment 嘅 model。

現有 model 包括 pay-per-call,

即係每個 agent invocation charge fixed fee,

例如 0.001 美金。

Subscription,

即係 monthly fee 包括 unlimited call 直至 quota。

Revenue share,

即係 marketplace platform take percentage 例如 30%,

開發者 take 70%。

Micropayment via crypto,

即係用 stablecoin 或者 L2 blockchain 做小額支付,

適合 high-frequency agent interaction。

[10 | 04:40] 嘉賓 F (expert 曉晴):

Turn 10 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 04:40 · section 2 (A2A Protocol & Agent Payment)

speaker=F · chars=315 · ts=04:40 · terms=agent-friendly, pay-per-call, micropayment, monetization, usage-based · sentences=5 · clauses=7

Verbatim phrases in this turn: Stripe 同 Coinbase 2025 年開始提供 agent-friendly API。 · lling 同 usage-based pricing, 適合 pay-per-call agent。 · L2 提供 cheap on-chain transaction, 適合 micropayment。

Stripe 同 Coinbase 2025 年開始提供 agent-friendly API。

First clause (verbatim): Stripe 同 Coinbase 2025 年開始提供 agent-friendly API。

Last clause (verbatim): Vercel 同 Netlify 提供 edge function deployment 適合 serverless agent。

Stripe 同 Coinbase 2025 年開始提供 agent-friendly API。

Stripe 嘅 metered billing 同 usage-based pricing,

適合 pay-per-call agent。

Coinbase 嘅 Base L2 提供 cheap on-chain transaction,

適合 micropayment。

Cloudflare 嘅 Agents SDK 提供 agent hosting 同 monetization framework。

Vercel 同 Netlify 提供 edge function deployment 適合 serverless agent。

End-of-section recap (last spoken sentence of A2A Protocol & Agent Payment): Stripe 同 Coinbase 2025 年開始提供 agent-friendly API。

Section 3/5 — Regulation & Economics

Regulation 同 Economics

Section overview: covers turns 11–16 (6 spoken segments).

Topic terms (extracted from spoken text): self-regulation, terminal-first, qualification, documentation, long-running, autocomplete, productivity, availability, unacceptable, transparency

Latin/English code-terms in this section (verbatim from speech): self-regulation, terminal-first, qualification

Section character total: 2,361 characters across 6 spoken turns.

Section duration estimate: ~3:06 of 14:00 total.

Turns in this section: 11, 12, 13, 14, 15, 16.

First spoken sentence of this section (turn 11, verbatim): Agent 嘅 use case 演進。

Average characters per turn (this section): ~393 chars.

Cumulative characters through this section: 5,639 of 9,528 total.

[11 | 05:11] 主持 M (host 子謙):

Turn 11 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 05:11 · section 3 (Regulation & Economics)

speaker=M · chars=464 · ts=05:11 · terms=long-running, multi-agent, automation, autonomous, persistent · sentences=5 · clauses=23

Verbatim phrases in this turn: e chatbot 加 tool use, 例如 ChatGPT Plugins、 · orkflow automation, 例如 AutoGen GroupChat、 · ing autonomous agent, 例如 OpenAI Operator、

, 2023 至 2024, 主要係 simple chatbot 加 tool use, 例如 ChatGPT Plugins、Claude Tool Use。

First clause (verbatim): Agent 嘅 use case 演進。

Last clause (verbatim): multi-agent economy。

Agent 嘅 use case 演進。

第一代 agent,

2023 至 2024,

主要係 simple chatbot 加 tool use,

例如 ChatGPT Plugins、

Claude Tool Use。

第二代 agent,

2024 至 2025,

multi-agent system 同 workflow automation,

例如 AutoGen GroupChat、

LangGraph multi-agent。

第三代 agent,

2025 至 2026,

開始有 long-running autonomous agent,

例如 OpenAI Operator、

Claude Computer Use、

Manus、

Devin 嘅 coding agent。

第四代 agent,

預計 2026 至 2027,

fully autonomous agent 自己 manage 自己嘅 task 同 resource,

persistent memory,

multi-agent economy。

[12 | 05:42] 嘉賓 F (expert 曉晴):

Turn 12 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 05:42 · section 3 (Regulation & Economics)

speaker=F · chars=426 · ts=05:42 · terms=qualification, engineering, Autonomous, commercial, multi-step · sentences=5 · clauses=17

Verbatim phrases in this turn: Software engineering, 例如 Devin、 · 研究 agent, 例如 OpenAI Deep Research、 · Customer service agent, 例如 Sierra、

Autonomous agent 嘅 commercial use case。

First clause (verbatim): Autonomous agent 嘅 commercial use case。

Last clause (verbatim): 可以做 lead outreach 同 qualification。

Autonomous agent 嘅 commercial use case。

Software engineering,

例如 Devin、

Cognition SWE-2 嘅 coding agent,

可以 handle multi-step coding task。

研究 agent,

例如 OpenAI Deep Research、

Google Gemini Deep Research,

可以做 hours-long research task。

Customer service agent,

例如 Sierra、

Decagon 嘅 customer service agent,

可以 handle 80% 嘅 inquiry 而 human 只 handle 20% complex case。

Sales agent,

例如 11x、

Regie 嘅 SDR agent,

可以做 lead outreach 同 qualification。

[13 | 06:13] 主持 M (host 子謙):

Turn 13 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 06:13 · section 3 (Regulation & Economics)

speaker=M · chars=351 · ts=06:13 · terms=terminal-first, autocomplete, autonomous, experience, enterprise · sentences=4 · clauses=9

Verbatim phrases in this turn: Coding agent 嘅 evolution 對 developer workflow 嘅影響。 · 26 年嘅 autonomous coding agent 例如 Devin 同 SWE-Agent。 · experience, Anthropic 同 OpenAI 都推出 enterprise plan。

Coding agent 嘅 evolution 對 developer workflow 嘅影響。

First clause (verbatim): Coding agent 嘅 evolution 對 developer workflow 嘅影響。

Last clause (verbatim): 反映 coding agent 嘅商業 value。

Coding agent 嘅 evolution 對 developer workflow 嘅影響。

Coding agent 已經由 2023 年嘅 simple autocomplete,

例如 Copilot,

進化到 2026 年嘅 autonomous coding agent 例如 Devin 同 SWE-Agent。

Claude Code 同 Codex CLI 提供 terminal-first coding experience,

Anthropic 同 OpenAI 都推出 enterprise plan。

OpenAI 嘅 Codex ARR 月增長 20.8%,

Claude Code 嘅 ARR 達到 151 億美金,

反映 coding agent 嘅商業 value。

[14 | 06:44] 嘉賓 F (expert 曉晴):

Turn 14 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 06:44 · section 3 (Regulation & Economics)

speaker=F · chars=344 · ts=06:44 · terms=productivity, availability, throughput, inference, customer · sentences=5 · clauses=12

Verbatim phrases in this turn: Agent 嘅 cost 包括 inference cost、 · tool call cost、 · t 嘅 productivity gain 包括 task throughput、

Agent 嘅經濟學。Agent 嘅 cost 包括 inference cost、tool call cost、storage cost。

First clause (verbatim): Agent 嘅經濟學。

Last clause (verbatim): Agent ROI 嘅計算需要考慮 cost saving 同 productivity gain。

Agent 嘅經濟學。

Agent 嘅 cost 包括 inference cost、

tool call cost、

storage cost。

Inference cost 2024 至 2026 年大幅下降,

例如 GPT-6 Luna 嘅 input token 0.1 美金 per million token,

比 GPT-5.6 Terra 減半。

Agent 嘅 productivity gain 包括 task throughput、

quality、

availability,

例如 24-7 availability 對 customer service 有顯著 value。

Agent ROI 嘅計算需要考慮 cost saving 同 productivity gain。

[15 | 07:15] 主持 M (host 子謙):

Turn 15 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 07:15 · section 3 (Regulation & Economics)

speaker=M · chars=416 · ts=07:15 · terms=documentation, unacceptable, transparency, regulation, applicable · sentences=4 · clauses=15

Verbatim phrases in this turn: risk AI system 需要 conformity assessment、 · risk management、 · data governance、

Agent 嘅 regulation 方向。

First clause (verbatim): Agent 嘅 regulation 方向。

Last clause (verbatim): 需要 extensive documentation 同 audit。

Agent 嘅 regulation 方向。

EU AI Act 喺 2025 年 fully applicable,

將 AI system 分 risk level,

包括 unacceptable risk 例如 social scoring,

high risk 例如 employment 同 credit decision,

limited risk 例如 chatbot,

minimal risk 例如 spam filter。

High risk AI system 需要 conformity assessment、

risk management、

data governance、

transparency、

human oversight。

Agent 因為可以 take action,

通常 fall 入 high risk category,

需要 extensive documentation 同 audit。

[16 | 07:46] 嘉賓 F (expert 曉晴):

Turn 16 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 07:46 · section 3 (Regulation & Economics)

speaker=F · chars=360 · ts=07:46 · terms=self-regulation, regulation, fragmented, California, Management · sentences=4 · clauses=9

Verbatim phrases in this turn: , 唔同 federal agency 各自有 guidance, 例如 FTC、 · 有 guidance, 例如 FTC、SEC、EEOC 都 publish AI guideline。 · 例如 California SB 1047 雖然被 veto 但反映 state level 嘅關注。

agmented, 唔同 federal agency 各自有 guidance, 例如 FTC、SEC、EEOC 都 publish AI guideline。

First clause (verbatim): US 嘅 AI regulation 比較 fragmented,

Last clause (verbatim): 產業 self-regulation 例如 NIST AI Risk Management Framework 同 Frontier Model Forum 嘅 best practice 暫時填補 federal 立法空缺。

US 嘅 AI regulation 比較 fragmented,

唔同 federal agency 各自有 guidance,

例如 FTC、

SEC、

EEOC 都 publish AI guideline。

州 level 立法例如 California SB 1047 雖然被 veto 但反映 state level 嘅關注。

Biden 嘅 Executive Order 14110 提供 broad framework,

但 Trump 2025 年上任後部分政策 reversal。

產業 self-regulation 例如 NIST AI Risk Management Framework 同 Frontier Model Forum 嘅 best practice 暫時填補 federal 立法空缺。

End-of-section recap (last spoken sentence of Regulation & Economics): US 嘅 AI regulation 比較 fragmented, 唔同 federal agency 各自有 guidance, 例如 FTC、SEC、EEOC 都 publish AI guideline。

Section 4/5 — Future Trends & Skill Path

Future Trends 同 Skill Path

Section overview: covers turns 17–22 (6 spoken segments).

Topic terms (extracted from spoken text): Superintelligence, production-grade, Recommendation, Administration, identification, Observability, OpenTelemetry, fundamentals, requirement, description

Latin/English code-terms in this section (verbatim from speech): Superintelligence, production-grade, Recommendation

Section character total: 2,428 characters across 6 spoken turns.

Section duration estimate: ~3:06 of 14:00 total.

Turns in this section: 17, 18, 19, 20, 21, 22.

First spoken sentence of this section (turn 17, verbatim): 中國 AI regulation。

Average characters per turn (this section): ~404 chars.

Cumulative characters through this section: 8,067 of 9,528 total.

[17 | 08:17] 主持 M (host 子謙):

Turn 17 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 08:17 · section 4 (Future Trends & Skill Path)

speaker=M · chars=367 · ts=08:17 · terms=Recommendation, Administration, requirement, description, regulation · sentences=5 · clauses=7

Verbatim phrases in this turn: 中國 AI regulation。 · service 做 security assessment 同 content moderation。 · eep Synthesis Provisions 即深度合成 規定 deepfake 等需要顯式標識。

es 2023 年生效, 要求 generative AI service 做 security assessment 同 content moderation。

First clause (verbatim): 中國 AI regulation。

Last clause (verbatim): 服務提供者需要 file algorithm 同 training data 嘅 description。

中國 AI regulation。

生成式 AI Service Management Measures 2023 年生效,

要求 generative AI service 做 security assessment 同 content moderation。

Deep Synthesis Provisions 即深度合成 規定 deepfake 等需要顯式標識。

Algorithm Recommendation Provisions 規定 algorithm recommendation 需要 opt-out 同時確保內容多元。

Cyberspace Administration 嘅 filing requirement,

服務提供者需要 file algorithm 同 training data 嘅 description。

[18 | 08:48] 嘉賓 F (expert 曉晴):

Turn 18 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 08:48 · section 4 (Future Trends & Skill Path)

speaker=F · chars=422 · ts=08:48 · terms=light-touch, regulation, management, governance, Securities · sentences=6 · clauses=10

Verbatim phrases in this turn: 香港 AI regulation。 · (Privacy) Ordinance 適用於 AI 處理 personal data 嘅 case。 · 有 guidance, 包括 model risk management 同 governance。

比較 light-touch, Personal Data (Privacy) Ordinance 適用於 AI 處理 personal data 嘅 case。

First clause (verbatim): 香港 AI regulation。

Last clause (verbatim): 但 increasing trend toward 監管。

香港 AI regulation。

香港嘅 AI policy 比較 light-touch,

Personal Data (Privacy) Ordinance 適用於 AI 處理 personal data 嘅 case。

Hong Kong Monetary Authority 對銀行使用 AI 有 guidance,

包括 model risk management 同 governance。

Securities and Futures Commission 對 AI 同 algo trading 有 specific guideline。

Innovation 同 Technology Bureau 推廣 AI adoption 同時 publish ethical AI framework。

整體而言香港係 friendlier 比 mainland,

比 EU 寬鬆,

但 increasing trend toward 監管。

[19 | 09:20] 主持 M (host 子謙):

Turn 19 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 09:20 · section 4 (Future Trends & Skill Path)

speaker=M · chars=317 · ts=09:20 · terms=Superintelligence, expectation, Responsible, definition, Anthropic · sentences=5 · clauses=8

Verbatim phrases in this turn: AGI 嘅 timeline 同 expectation。 · 25 年公開呼籲放慢 AI 開發步伐加強安全同保安測試, 呢個 joint appeal 係業界罕見。 · 同 Responsible Scaling Policy 提供 internal framework。

AGI 嘅 timeline 同 expectation。

First clause (verbatim): AGI 嘅 timeline 同 expectation。

Last clause (verbatim): 主張 world model approach。

AGI 嘅 timeline 同 expectation。

Sam Altman 同 Dario Amodei 喺 2025 年公開呼籲放慢 AI 開發步伐加強安全同保安測試,

呢個 joint appeal 係業界罕見。

OpenAI 同 Anthropic 嘅 Superintelligence 同 Responsible Scaling Policy 提供 internal framework。

Demis Hassabis 預測 AGI 可能 5 至 10 年內出現,

但 definition 唔統一。

Yann LeCun 對 LLM-based AGI path 表示懷疑,

主張 world model approach。

[20 | 09:51] 嘉賓 F (expert 曉晴):

Turn 20 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 09:51 · section 4 (Future Trends & Skill Path)

speaker=F · chars=401 · ts=09:51 · terms=participate, multimodal, processing, scientific, prediction · sentences=5 · clauses=15

Verbatim phrases in this turn: multimodal agent, 即係 LLM 不只 text 而係 text、 · nt, 即係 agent 控制 physical robot 例如 Figure、 · Tesla Optimus、

Agent 嘅 future trend。

First clause (verbatim): Agent 嘅 future trend。

Last clause (verbatim): 即係 agent 自己 manage 自己嘅 resource 同 participate 喺 economy 入面。

Agent 嘅 future trend。

第一個係 multimodal agent,

即係 LLM 不只 text 而係 text、

image、

audio、

video unified processing,

例如 GPT-6 Astra 同 Claude Opus 5.5。

第二個係 embodied agent,

即係 agent 控制 physical robot 例如 Figure、

Tesla Optimus、

Unitree。

第三個係 scientific agent,

例如 AlphaFold 嘅 protein structure prediction 同材料 science 嘅 GNoME 加速 discovery 過程。

第四個係 economic agent,

即係 agent 自己 manage 自己嘅 resource 同 participate 喺 economy 入面。

[21 | 10:22] 主持 M (host 子謙):

Turn 21 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 10:22 · section 4 (Future Trends & Skill Path)

speaker=M · chars=490 · ts=10:22 · terms=production-grade, identification, Observability, OpenTelemetry, fundamentals · sentences=9 · clauses=20

Verbatim phrases in this turn: LLM fundamentals, 包括 prompting、 · fine-tuning、 · Tool integration 包括 API、

Agent development 嘅 skill path。

First clause (verbatim): Agent development 嘅 skill path。

Last clause (verbatim): Business acumen 包括 use case identification 同 ROI calculation。

Agent development 嘅 skill path。

如果你想 build production agent,

recommended skill path 包括 Python proficiency,

因為大部分 framework 喺 Python ecosystem。

LLM fundamentals,

包括 prompting、

fine-tuning、

evaluation。

One framework deep dive,

例如 LangGraph 因為 production-grade。

Tool integration 包括 API、

MCP、

database。

Deployment 包括 Docker、

Kubernetes、

CI/CD。

Observability 包括 LangSmith、

OpenTelemetry。

Safety 同 alignment 包括 prompt injection defense 同 sandboxing。

Business acumen 包括 use case identification 同 ROI calculation。

[22 | 10:53] 嘉賓 F (expert 曉晴):

Turn 22 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 10:53 · section 4 (Future Trends & Skill Path)

speaker=F · chars=431 · ts=10:53 · terms=development, improvement, Marketplace, production, Researcher · sentences=6 · clauses=14

Verbatim phrases in this turn: Agent development 嘅 career path。 · ld production agent 嘅 engineer, 2026 年 demand 急速上升。 · pattern 同 improvement, 通常喺 academic 同 frontier lab。

Agent development 嘅 career path。

First clause (verbatim): Agent development 嘅 career path。

Last clause (verbatim): Agent Economy 嘅新角色包括 Agent Marketplace Operator 同 Agent Reputation Service Provider。

Agent development 嘅 career path。

Agent Engineer,

即係 build production agent 嘅 engineer,

2026 年 demand 急速上升。

Agent Researcher,

即係研究新 pattern 同 improvement,

通常喺 academic 同 frontier lab。

Agent PM,

即係 identify use case 同 define product roadmap,

需要 AI 同 business 雙重 acumen。

Agent Safety Engineer,

即係 focus 喺 alignment 同 security,

隨着 regulation 收緊 demand 上升。

Agent Economy 嘅新角色包括 Agent Marketplace Operator 同 Agent Reputation Service Provider。

End-of-section recap (last spoken sentence of Future Trends & Skill Path): Agent development 嘅 career path。

Section 5/5 — Course Wrap-up

課程總結

Section overview: covers turns 23–27 (5 spoken segments).

Topic terms (extracted from spoken text): Supervisor-Worker, Plan-and-Execute, multi-subsystem, implementation, multi-language, infrastructure, reversibility, architectural, documentation, Architecture

Latin/English code-terms in this section (verbatim from speech): Supervisor-Worker, Plan-and-Execute, multi-subsystem

Section character total: 1,461 characters across 5 spoken turns.

Section duration estimate: ~2:35 of 14:00 total.

Turns in this section: 23, 24, 25, 26, 27.

First spoken sentence of this section (turn 23, verbatim): 課程總結。

Average characters per turn (this section): ~292 chars.

Cumulative characters through this section: 9,528 of 9,528 total.

[23 | 11:24] 主持 M (host 子謙):

Turn 23 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 11:24 · section 5 (Course Wrap-up)

speaker=M · chars=380 · ts=11:24 · terms=Supervisor-Worker, Plan-and-Execute, implementation, Architecture, per-pattern · sentences=10 · clauses=19

Verbatim phrases in this turn: 第一堂講五大架構 pattern 包括 ReAct、 · Plan-and-Execute、 · Reflexion、

課程總結。我哋用咗八堂由 abstract 講到 implementation。

First clause (verbatim): 課程總結。

Last clause (verbatim): 第八堂講 Economy 同 Future。

課程總結。

我哋用咗八堂由 abstract 講到 implementation。

第一堂講五大架構 pattern 包括 ReAct、

Plan-and-Execute、

Reflexion、

Supervisor-Worker、

Workflow Graph。

第二堂講四大 framework 包括 LangGraph、

CrewAI、

MAF、

ADK。

第三堂講 Memory Architecture 包括 episodic、

semantic、

procedural memory 同 RAG。

第四堂講 Tool Use 同 MCP。

第五堂講 per-pattern implementation walkthrough。

第六堂講 Evaluation 同 Benchmarks。

第七堂講 Safety 同 Alignment。

第八堂講 Economy 同 Future。

[24 | 11:55] 嘉賓 F (expert 曉晴):

Turn 24 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 11:55 · section 5 (Course Wrap-up)

speaker=F · chars=430 · ts=11:55 · terms=multi-subsystem, multi-language, reversibility, architecture, engineering · sentences=5 · clauses=16

Verbatim phrases in this turn: 揀 pattern base on task duration、 · action reversibility、 · se case, LangGraph 對 production stateful、

Final 嘅 take-away。Agent 係 control flow 嘅 architecture, 唔係純粹 prompt engineering。

First clause (verbatim): Final 嘅 take-away。

Last clause (verbatim): episodic 加 semantic 加 procedural 唔係 single store。

Final 嘅 take-away。

Agent 係 control flow 嘅 architecture,

唔係純粹 prompt engineering。

揀 pattern base on task duration、

action reversibility、

autonomy level,

由最簡單開始,

失敗 mode 出現之後先升 complexity。

Framework 係 tool,

揀 framework base on use case,

LangGraph 對 production stateful、

CrewAI 對 fast role-based prototype、

MAF 對 Microsoft stack、

ADK 對 Google Cloud 同 multi-language。

Memory 係 multi-subsystem,

episodic 加 semantic 加 procedural 唔係 single store。

[25 | 12:26] 主持 M (host 子謙):

Turn 25 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 12:26 · section 5 (Course Wrap-up)

speaker=M · chars=378 · ts=12:26 · terms=infrastructure, architectural, documentation, afterthought, prompt-level · sentences=3 · clauses=11

Verbatim phrases in this turn: tural layer 嘅 defense 包括 least privilege、 · sandboxing、 · approval gate、

ge、sandboxing、approval gate、prompt injection defense, prompt-level guardrail 唔足夠。

First clause (verbatim): Safety 唔可以 afterthought,

Last clause (verbatim): EU AI Act 同各地 guideline 將令 high-risk agent 需要 extensive documentation 同 audit。

Safety 唔可以 afterthought,

必須 architectural layer 嘅 defense 包括 least privilege、

sandboxing、

approval gate、

prompt injection defense,

prompt-level guardrail 唔足夠。

Economy 會 evolve,

A2A protocol 同 marketplace 將令 agent 之間可以 discover 同 transact,

identity 同 reputation 變成 critical infrastructure。

Regulation 會加強,

EU AI Act 同各地 guideline 將令 high-risk agent 需要 extensive documentation 同 audit。

[26 | 12:57] 嘉賓 F (expert 曉晴):

Turn 26 of 27 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 12:57 · section 5 (Course Wrap-up)

speaker=F · chars=260 · ts=12:57 · terms=responsible, development, powerful, Telegram, purpose · sentences=4 · clauses=8

Verbatim phrases in this turn: gent 嘅 future 係 bright 但需要 responsible development。 · ent, 而係教你點 build 一個 powerful 同 responsible 嘅 agent。 · 謝各位同學八堂嘅參與, 我哋希望呢個 course 對你嘅 AI agent journey 有幫助。

最後, agent 嘅 future 係 bright 但需要 responsible development。

First clause (verbatim): 最後,

Last clause (verbatim): 可以通過 podcast channel 或者 Telegram thread 同我哋聯繫。

最後,

agent 嘅 future 係 bright 但需要 responsible development。

我哋呢個 course 嘅 purpose 唔係教你點 build 一個 powerful agent,

而係教你點 build 一個 powerful 同 responsible 嘅 agent。

多謝各位同學八堂嘅參與,

我哋希望呢個 course 對你嘅 AI agent journey 有幫助。

如果有問題,

可以通過 podcast channel 或者 Telegram thread 同我哋聯繫。

[27 | 13:28] 主持 M (host 子謙):

Turn 27 of 27 · speaker M (host 子謙 — opens and closes) · audio timestamp 13:28 · section 5 (Course Wrap-up)

speaker=M · chars=13 · ts=13:28 · sentences=1 · clauses=2

First clause (verbatim): 我哋下次再見,

Last clause (verbatim): 各位同學。

我哋下次再見,

各位同學。

End-of-section recap (last spoken sentence of Course Wrap-up): 我哋下次再見, 各位同學。

End-of-lesson summary

This lesson covered 5 sections across 27 spoken turns (~14 min audio). Below is the final sentence of each section, preserved verbatim from the source podcast script.

• Opening & Agent Marketplace (turn 04): 現有嘅 Agent Marketplace 例子。
• A2A Protocol & Agent Payment (turn 10): Stripe 同 Coinbase 2025 年開始提供 agent-friendly API。
• Regulation & Economics (turn 16): US 嘅 AI regulation 比較 fragmented, 唔同 federal agency 各自有 guidance, 例如 FTC、SEC、EEOC 都 publish AI guideline。
• Future Trends & Skill Path (turn 22): Agent development 嘅 career path。
• Course Wrap-up (turn 27): 我哋下次再見, 各位同學。

End of transcript

Total turns in this lesson: 27 spoken segments · ~14 min audio · preserved verbatim from the source podcast script (/opt/data/workspace/projects/ai-agent-course-08/script_raw.json).

Use the audio player above to listen along. The Quiz section below tests comprehension of this lesson.

Source & integrity

• Source file: script_raw.json (the line-by-line Cantonese dialogue that was TTS-synthesised into the lesson MP3)
• Fidelity: all 廣東話 text is byte-identical to the source — no translation, no summarisation, no paraphrasing
• Markdown structure added: speaker labels, section headings, timestamp markers, per-turn metadata callouts (speaker id, char count, timestamp, verbatim terms), per-section overview blocks, sentence-level line breaks (for readability only)
• Rendering: react-markdown + remark-gfm in TranscriptPanel.tsx; dark theme & mobile-responsive via Tailwind prose

Integrity checksum

• Total spoken characters (across all turns): 9,528
• Total spoken sentences (across all turns): 132
• Total spoken clauses (across all turns): 322
• Speaker turn distribution: M=14 · F=13