Lesson 5: Per-Pattern Implementation Walkthrough — 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 5 of 8 · learnagent.lmmlab.com Topic (EN): Python code for ReAct, Plan-and-Execute, Reflexion, Supervisor-Worker, Workflow Graph. Topic (粵): 五大 pattern 嘅 Python 實作。 Speakers: 主持 M (host 子謙) and 嘉賓 F (expert 曉晴) · 25 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 & ReAct Implementation (粵: 開場同 ReAct 實作) — turn 01 onwards
• 2. Plan-and-Execute & Reflexion Implementation (粵: Plan-and-Execute 同 Reflexion 實作) — turn 05 onwards
• 3. Supervisor-Worker & Workflow Graph Implementation (粵: Supervisor-Worker 同 Workflow Graph 實作) — turn 11 onwards
• 4. Production Considerations (粵: Production 考量) — turn 17 onwards
• 5. Wrap-up & Evaluation Preview (粵: 總結同 Evaluation 預覽) — turn 23 onwards

Section 1/5 — Opening & ReAct Implementation

開場同 ReAct 實作

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

Topic terms (extracted from spoken text): create_react_agent, langgraph.prebuilt, Supervisor-Worker, Plan-and-Execute, init_chat_model, Implementation, max_iterations, token-by-token, SystemMessage, HumanMessage

Latin/English code-terms in this section (verbatim from speech): create_react_agent, langgraph.prebuilt, Supervisor-Worker

Section character total: 994 characters across 4 spoken turns.

Section duration estimate: ~2:14 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): ~248 chars.

Cumulative characters through this section: 994 of 8,216 total.

[01 | 00:00] 主持 M (host 子謙):

Turn 1 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 00:00 · section 1 (Opening & ReAct Implementation)

speaker=M · chars=199 · ts=00:00 · terms=Supervisor-Worker, Plan-and-Execute, Implementation, Per-Pattern, Walkthrough · sentences=3 · clauses=9

Verbatim phrases in this turn: tion Walkthrough, 用 Python code 示範 ReAct、 · Plan-and-Execute、 · Reflexion、

各位同學早晨, 我係子謙。歡迎收聽第五課。

First clause (verbatim): 各位同學早晨,

Last clause (verbatim): 主要用 LangGraph 同 LangChain 嘅 API。

各位同學早晨,

我係子謙。

歡迎收聽第五課。

今日嘅主題係 Per-Pattern Implementation Walkthrough,

用 Python code 示範 ReAct、

Plan-and-Execute、

Reflexion、

Supervisor-Worker 四個 pattern 嘅具體 implementation,

主要用 LangGraph 同 LangChain 嘅 API。

[02 | 00:33] 嘉賓 F (expert 曉晴):

Turn 2 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 00:33 · section 1 (Opening & ReAct Implementation)

speaker=F · chars=109 · ts=00:33 · terms=contract, pattern, combine, project, tool · sentences=3 · clauses=6

Verbatim phrases in this turn: 第一堂講咗 pattern 嘅抽象理論, 第四課講咗 tool use 嘅 contract。 · 日就將兩個 combine 落具體 code, 令你可以直接 copy 落自己嘅 project 用。

大家好, 我係曉晴。第一堂講咗 pattern 嘅抽象理論, 第四課講咗 tool use 嘅 contract。

First clause (verbatim): 大家好,

Last clause (verbatim): 令你可以直接 copy 落自己嘅 project 用。

大家好,

我係曉晴。

第一堂講咗 pattern 嘅抽象理論,

第四課講咗 tool use 嘅 contract。

今日就將兩個 combine 落具體 code,

令你可以直接 copy 落自己嘅 project 用。

[03 | 01:07] 主持 M (host 子謙):

Turn 3 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 01:07 · section 1 (Opening & ReAct Implementation)

speaker=M · chars=325 · ts=01:07 · terms=create_react_agent, langgraph.prebuilt, init_chat_model, implementation, identifier · sentences=4 · clauses=11

Verbatim phrases in this turn: prebuilt 引入 create_react_agent, 然後 model、 · 首先講 ReAct 嘅 LangGraph implementation。 · react_agent 嘅 prebuilt function, 係最簡單嘅 entry point。

首先講 ReAct 嘅 LangGraph implementation。

First clause (verbatim): 首先講 ReAct 嘅 LangGraph implementation。

Last clause (verbatim): tools 等於 list of tool function。

首先講 ReAct 嘅 LangGraph implementation。

LangGraph 提供 create_react_agent 嘅 prebuilt function,

係最簡單嘅 entry point。

基本 skeleton 係,

由 langgraph.prebuilt 引入 create_react_agent,

然後 model、

tools、

prompt 傳入去。

即係 from langgraph.prebuilt import create_react_agent,

model 等於 init_chat_model 或者 string identifier,

tools 等於 list of tool function。

[04 | 01:40] 嘉賓 F (expert 曉晴):

Turn 4 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 01:40 · section 1 (Opening & ReAct Implementation)

speaker=F · chars=361 · ts=01:40 · terms=max_iterations, token-by-token, SystemMessage, HumanMessage, agent.stream · sentences=4 · clauses=16

Verbatim phrases in this turn: 係 list of message dicts, 例如 HumanMessage、 · SystemMessage、 · AIMessage、

ReAct agent 嘅 invoke 嘅 usage。

First clause (verbatim): ReAct agent 嘅 invoke 嘅 usage。

Last clause (verbatim): 包括 tool call start 同 finish。

ReAct agent 嘅 invoke 嘅 usage。

Agent 接受 messages,

係 list of message dicts,

例如 HumanMessage、

SystemMessage、

AIMessage、

ToolMessage。

當 invoke 嘅時候,

agent 會 iterate thought、

action、

observation loop,

直至 task done 或者 hit max_iterations。

如果要 stream token-by-token,

可以用 agent.stream 嘅 mode values,

mode values 嘅 stream mode 會 yield 每個 state change,

包括 tool call start 同 finish。

End-of-section recap (last spoken sentence of Opening & ReAct Implementation): ReAct agent 嘅 invoke 嘅 usage。

Section 2/5 — Plan-and-Execute & Reflexion Implementation

Plan-and-Execute 同 Reflexion 實作

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

Topic terms (extracted from spoken text): PostgresSaver.from_conn_string, langgraph.checkpoint.postgres, multi-conversation, Human-in-the-loop, interrupt_before, Plan-and-Execute, straightforward, implementation, customization, PostgresSaver

Latin/English code-terms in this section (verbatim from speech): PostgresSaver.from_conn_string, langgraph.checkpoint.postgres, multi-conversation

Section character total: 2,037 characters across 6 spoken turns.

Section duration estimate: ~3:21 of 14:00 total.

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

First spoken sentence of this section (turn 05, verbatim): ReAct agent 嘅 customization。

Average characters per turn (this section): ~339 chars.

Cumulative characters through this section: 3,031 of 8,216 total.

[05 | 02:14] 主持 M (host 子謙):

Turn 5 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 02:14 · section 2 (Plan-and-Execute & Reflexion Implementation)

speaker=M · chars=390 · ts=02:14 · terms=Human-in-the-loop, interrupt_before, customization, checkpointer, conversation · sentences=5 · clauses=14

Verbatim phrases in this turn: ReAct agent 嘅 customization。 · er 做 conversation memory, 跨 invoke 嘅 state persist。 · es, 例如 trim conversation history 或者 inject context。

ReAct agent 嘅 customization。

First clause (verbatim): ReAct agent 嘅 customization。

Last clause (verbatim): 等 human approve。

ReAct agent 嘅 customization。

Memory,

可以加 checkpointer 例如 MemorySaver 做 conversation memory,

跨 invoke 嘅 state persist。

Pre-model hook,

可以加一個 node 喺 model call 之前 modify messages,

例如 trim conversation history 或者 inject context。

Post-model hook,

加一個 node 喺 model call 之後,

例如 log 或者 trigger validation。

Human-in-the-loop,

加 interrupt_before 喺 sensitive tool,

即係 tool call 之前 pause,

等 human approve。

[06 | 02:48] 嘉賓 F (expert 曉晴):

Turn 6 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 02:48 · section 2 (Plan-and-Execute & Reflexion Implementation)

speaker=F · chars=363 · ts=02:48 · terms=PostgresSaver.from_conn_string, langgraph.checkpoint.postgres, multi-conversation, PostgresSaver, checkpoint_ns · sentences=3 · clauses=7

Verbatim phrases in this turn: 因為 MemorySaver 係 in-memory, restart 之後會 loss state。 · gresSaver.from_conn_string 用 connection string 初始化。 · ns, 支援 multi-tenant 同 multi-conversation isolation。

checkpointer 做 durable state, 因為 MemorySaver 係 in-memory, restart 之後會 loss state。

First clause (verbatim): ReAct 嘅 production deployment 通常配 PostgreSQL checkpointer 做 durable state,

Last clause (verbatim): 支援 multi-tenant 同 multi-conversation isolation。

ReAct 嘅 production deployment 通常配 PostgreSQL checkpointer 做 durable state,

因為 MemorySaver 係 in-memory,

restart 之後會 loss state。

PostgreSQL checkpoint 嘅 setup 係 from langgraph.checkpoint.postgres import PostgresSaver,

然後 PostgresSaver.from_conn_string 用 connection string 初始化。

Checkpointer 接受 thread_id 同 checkpoint_ns,

支援 multi-tenant 同 multi-conversation isolation。

[07 | 03:21] 主持 M (host 子謙):

Turn 7 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 03:21 · section 2 (Plan-and-Execute & Reflexion Implementation)

speaker=M · chars=359 · ts=03:21 · terms=Plan-and-Execute, implementation, observation, LangGraph, replanner · sentences=6 · clauses=14

Verbatim phrases in this turn: Plan-and-Execute 通常有 planner、 · executor、 · 好, 第二個 pattern, Plan-and-Execute 嘅 implementation。

好, 第二個 pattern, Plan-and-Execute 嘅 implementation。

First clause (verbatim): 好,

Last clause (verbatim): 決定繼續執行或者重新 plan。

好,

第二個 pattern,

Plan-and-Execute 嘅 implementation。

LangGraph 唔似 ReAct 有 prebuilt function,

所以需要 explicit 寫 graph。

Plan-and-Execute 通常有 planner、

executor、

replanner 三個 nodes。

Planner 接受 messages 同 plan state,

output 一個 ordered list of steps。

Executor 接受 current step 同 tools,

執行 step 同 return observation。

Replanner 接受 completed steps 同 failed steps,

決定繼續執行或者重新 plan。

[08 | 03:55] 嘉賓 F (expert 曉晴):

Turn 8 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 03:55 · section 2 (Plan-and-Execute & Reflexion Implementation)

speaker=F · chars=279 · ts=03:55 · terms=Plan-and-Execute, conditional, past_steps, TypedDict, replanner · sentences=6 · clauses=10

Verbatim phrases in this turn: State 入面有 messages、 · past_steps、 · Plan-and-Execute 嘅 state 通常用 TypedDict 定義。

Plan-and-Execute 嘅 state 通常用 TypedDict 定義。

First clause (verbatim): Plan-and-Execute 嘅 state 通常用 TypedDict 定義。

Last clause (verbatim): Graph 嘅 conditional edge 由 replanner decide 去 end node 還是 back to executor。

Plan-and-Execute 嘅 state 通常用 TypedDict 定義。

State 入面有 messages、

plan、

past_steps、

response。

Plan 係 list of strings 或者 list of dicts。

Past steps 係 list of tuples,

每個 tuple 係 step 同 result。

Response 係 final answer。

Graph 嘅 conditional edge 由 replanner decide 去 end node 還是 back to executor。

[09 | 04:28] 主持 M (host 子謙):

Turn 9 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 04:28 · section 2 (Plan-and-Execute & Reflexion Implementation)

speaker=M · chars=294 · ts=04:28 · terms=implementation, past_steps, Replanner, messages, response · sentences=4 · clauses=14

Verbatim phrases in this turn: mplementation 通常係一個 LLM call 接受 messages、 · sages、plan、past_steps, output 一個 response 或者新 plan。 · 如果 plan 已經 complete, output response 同 end node。

entation 通常係一個 LLM call 接受 messages、plan、past_steps, output 一個 response 或者新 plan。

First clause (verbatim): Replanner 嘅 implementation 通常係一個 LLM call 接受 messages、

Last clause (verbatim): 或者 finish。

Replanner 嘅 implementation 通常係一個 LLM call 接受 messages、

plan、

past_steps,

output 一個 response 或者新 plan。

如果 plan 已經 complete,

output response 同 end node。

如果 step 失敗,

output revised plan 取代原來 plan。

Replanner prompt 要 explicit,

例如你係一個 replanner,

given 過去嘅 steps 同 results,

decide 要繼續,

重新 plan,

或者 finish。

[10 | 05:02] 嘉賓 F (expert 曉晴):

Turn 10 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 05:02 · section 2 (Plan-and-Execute & Reflexion Implementation)

speaker=F · chars=352 · ts=05:02 · terms=Plan-and-Execute, straightforward, optimization, granularity, competitor · sentences=3 · clauses=7

Verbatim phrases in this turn: Plan-and-Execute 嘅 cost optimization。 · depth, 但 execution 通常係 straightforward 嘅 tool call。 · y name 更加 robust 但需要 executor 自己 figure out detail。

Plan-and-Execute 嘅 cost optimization。

First clause (verbatim): Plan-and-Execute 嘅 cost optimization。

Last clause (verbatim): 高 level plan 例如 research competitor 比起低 level plan 例如 search company name 更加 robust 但需要 executor 自己 figure out detail。

Plan-and-Execute 嘅 cost optimization。

Planner 可以用 capable model 例如 Opus 5 或者 Sonnet 5,

Executor 可以用 cheap model 例如 Luna 或者 Mini,

因為 planning 需要 reasoning depth,

但 execution 通常係 straightforward 嘅 tool call。

Plan 嘅 granularity 要 control,

高 level plan 例如 research competitor 比起低 level plan 例如 search company name 更加 robust 但需要 executor 自己 figure out detail。

End-of-section recap (last spoken sentence of Plan-and-Execute & Reflexion Implementation): Plan-and-Execute 嘅 cost optimization。

Section 3/5 — Supervisor-Worker & Workflow Graph Implementation

Supervisor-Worker 同 Workflow Graph 實作

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

Topic terms (extracted from spoken text): Supervisor-Worker, Plan-and-Execute, Critic-augmented, implementation, research_agent, reviewer_agent, self-critique, sanitization, writer_agent, improvement

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

Section character total: 2,005 characters across 6 spoken turns.

Section duration estimate: ~3:21 of 14:00 total.

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

First spoken sentence of this section (turn 11, verbatim): 好, 第三個 pattern, Reflexion 嘅 implementation。

Average characters per turn (this section): ~334 chars.

Cumulative characters through this section: 5,036 of 8,216 total.

[11 | 05:36] 主持 M (host 子謙):

Turn 11 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 05:36 · section 3 (Supervisor-Worker & Workflow Graph Implementation)

speaker=M · chars=243 · ts=05:36 · terms=Plan-and-Execute, implementation, reflection, Reflexion, pattern · sentences=3 · clauses=10

Verbatim phrases in this turn: 好, 第三個 pattern, Reflexion 嘅 implementation。 · eAct 或者 Plan-and-Execute 做底, 外面再加 reflection layer。 · ct reflection, 再 retry with reflection 喺 prompt 入面。

好, 第三個 pattern, Reflexion 嘅 implementation。

First clause (verbatim): 好,

Last clause (verbatim): 再 retry with reflection 喺 prompt 入面。

好,

第三個 pattern,

Reflexion 嘅 implementation。

Reflexion 通常 wrap 其他 pattern,

即係 ReAct 或者 Plan-and-Execute 做底,

外面再加 reflection layer。

最簡單嘅 implementation 係做 retry loop,

agent 第一次 attempt,

失敗嘅話 collect reflection,

再 retry with reflection 喺 prompt 入面。

[12 | 06:09] 嘉賓 F (expert 曉晴):

Turn 12 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 06:09 · section 3 (Supervisor-Worker & Workflow Graph Implementation)

speaker=F · chars=326 · ts=06:09 · terms=self-critique, improvement, reflection, suggestion, Reflexion · sentences=3 · clauses=9

Verbatim phrases in this turn: 己嘅 previous attempt, output improvement suggestion。 · eas for improvement, write 2-3 sentence reflection。 · ory buffer, 下次 attempt 嘅 prompt prepend reflection。

可以係 self-critique, 即係 LLM 評估自己嘅 previous attempt, output improvement suggestion。

First clause (verbatim): Reflexion 嘅 reflection generator 可以係 self-critique,

Last clause (verbatim): 下次 attempt 嘅 prompt prepend reflection。

Reflexion 嘅 reflection generator 可以係 self-critique,

即係 LLM 評估自己嘅 previous attempt,

output improvement suggestion。

Reflection prompt 要 explicit,

例如 review 你剛才嘅 response,

identify mistakes 同 areas for improvement,

write 2-3 sentence reflection。

Reflection 可以 store 喺 episodic memory buffer,

下次 attempt 嘅 prompt prepend reflection。

[13 | 06:43] 主持 M (host 子謙):

Turn 13 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 06:43 · section 3 (Supervisor-Worker & Workflow Graph Implementation)

speaker=M · chars=372 · ts=06:43 · terms=Critic-augmented, Multi-trial, reflection, Reflexion, objective · sentences=5 · clauses=11

Verbatim phrases in this turn: Reflexion 嘅 advanced pattern。 · ect 多次, 例如每次 reflection 都 build 喺上一次 reflection 上面。 · primary agent 嘅 output, 提供 more objective critique。

Reflexion 嘅 advanced pattern。

First clause (verbatim): Reflexion 嘅 advanced pattern。

Last clause (verbatim): 適合有 label 嘅 task。

Reflexion 嘅 advanced pattern。

Multi-trial reflection,

即係 reflect 多次,

例如每次 reflection 都 build 喺上一次 reflection 上面。

Critic-augmented reflection,

即係另一個 agent 或者 tool 評估 primary agent 嘅 output,

提供 more objective critique。

例如 unit test failure message 可以自動做 reflection source。

External evaluator reflection,

即係用 ground truth 或者 expert feedback 做 reflection source,

適合有 label 嘅 task。

[14 | 07:16] 嘉賓 F (expert 曉晴):

Turn 14 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 07:16 · section 3 (Supervisor-Worker & Workflow Graph Implementation)

speaker=F · chars=362 · ts=07:16 · terms=sanitization, aggregation, reflections, concatenate, escalation · sentences=5 · clauses=12

Verbatim phrases in this turn: Reflexion 嘅 best practice。 · etry cap, 例如最多 retry 三次, 否則 fallback to escalation。 · e add 去 memory, 因為 malformed reflection 會污染 memory。

Reflexion 嘅 best practice。Retry cap, 例如最多 retry 三次, 否則 fallback to escalation。

First clause (verbatim): Reflexion 嘅 best practice。

Last clause (verbatim): 否則 prompt 會 redundant 同 contradict。

Reflexion 嘅 best practice。

Retry cap,

例如最多 retry 三次,

否則 fallback to escalation。

Reflection sanitization,

即係 reflection entry 要 verified before add 去 memory,

因為 malformed reflection 會污染 memory。

Reflection expiry,

即係 old reflection 要 evict 因為 context 可能 stale。

Reflection aggregation,

即係多個 reflections 要 synthesize 唔係簡單 concatenate,

否則 prompt 會 redundant 同 contradict。

[15 | 07:50] 主持 M (host 子謙):

Turn 15 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 07:50 · section 3 (Supervisor-Worker & Workflow Graph Implementation)

speaker=M · chars=373 · ts=07:50 · terms=Supervisor-Worker, implementation, research_agent, Supervisor, competitor · sentences=5 · clauses=12

Verbatim phrases in this turn: 好, 第四個 pattern, Supervisor-Worker 嘅 implementation。 · 通常係主 graph, 每個 Worker 係 subgraph 或者 function node。 · r is research_agent, task is research competitor X。

好, 第四個 pattern, Supervisor-Worker 嘅 implementation。

First clause (verbatim): 好,

Last clause (verbatim): Supervisor 收到 worker result 之後決定 next worker 或者 synthesize final answer。

好,

第四個 pattern,

Supervisor-Worker 嘅 implementation。

LangGraph 入面 Supervisor 通常係主 graph,

每個 Worker 係 subgraph 或者 function node。

Supervisor node 嘅 output 係 worker selection 同 task payload,

例如 next worker is research_agent,

task is research competitor X。

Worker node 接受 task payload,

執行 task,

return result。

Supervisor 收到 worker result 之後決定 next worker 或者 synthesize final answer。

[16 | 08:24] 嘉賓 F (expert 曉晴):

Turn 16 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 08:24 · section 3 (Supervisor-Worker & Workflow Graph Implementation)

speaker=F · chars=329 · ts=08:24 · terms=Supervisor-Worker, research_agent, reviewer_agent, writer_agent, conditional · sentences=4 · clauses=8

Verbatim phrases in this turn: rker 嘅 conditional edge 通常由 Supervisor 嘅 output 控制。 · h, writer_agent 做 writing, reviewer_agent 做 review。 · decide which worker to call next 或者 return FINISH。

Supervisor-Worker 嘅 conditional edge 通常由 Supervisor 嘅 output 控制。

First clause (verbatim): Supervisor-Worker 嘅 conditional edge 通常由 Supervisor 嘅 output 控制。

Last clause (verbatim): Supervisor 通常用 capable model 因為 routing decision critical。

Supervisor-Worker 嘅 conditional edge 通常由 Supervisor 嘅 output 控制。

Supervisor 嘅 prompt 要 define workers 同佢哋嘅 capability,

例如你有一個 research_agent 做 research,

writer_agent 做 writing,

reviewer_agent 做 review。

Given current state,

decide which worker to call next 或者 return FINISH。

Supervisor 通常用 capable model 因為 routing decision critical。

End-of-section recap (last spoken sentence of Supervisor-Worker & Workflow Graph Implementation): Supervisor-Worker 嘅 conditional edge 通常由 Supervisor 嘅 output 控制。

Section 4/5 — Production Considerations

Production 考量

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

Topic terms (extracted from spoken text): create_react_agent, Supervisor-Worker, research_agent, writer_agent, out-of-scope, restriction, independent, conditional, exponential, multi-agent

Latin/English code-terms in this section (verbatim from speech): create_react_agent, Supervisor-Worker, research_agent

Section character total: 2,096 characters across 6 spoken turns.

Section duration estimate: ~3:21 of 14:00 total.

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

First spoken sentence of this section (turn 17, verbatim): Worker 嘅 design。

Average characters per turn (this section): ~349 chars.

Cumulative characters through this section: 7,132 of 8,216 total.

[17 | 08:57] 主持 M (host 子謙):

Turn 17 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 08:57 · section 4 (Production Considerations)

speaker=M · chars=291 · ts=08:57 · terms=Supervisor, reasoning, subgraph, sub-task, contract · sentences=5 · clauses=10

Verbatim phrases in this turn: Worker 嘅 design。 · r 可以係獨立 agent 例如 ReAct agent, 適合需要 tool use 嘅 task。 · r 可以係 simple LLM call, 適合 pure reasoning 例如 review。

Worker 嘅 design。Worker 可以係獨立 agent 例如 ReAct agent, 適合需要 tool use 嘅 task。

First clause (verbatim): Worker 嘅 design。

Last clause (verbatim): Supervisor 要知道點樣 format input 同 parse output。

Worker 嘅 design。

Worker 可以係獨立 agent 例如 ReAct agent,

適合需要 tool use 嘅 task。

Worker 可以係 simple LLM call,

適合 pure reasoning 例如 review。

Worker 可以係 subgraph,

即係 Worker 本身有自己嘅 control flow,

適合 complex sub-task。

每個 Worker 嘅 input 同 output 要有 clear contract,

Supervisor 要知道點樣 format input 同 parse output。

[18 | 09:31] 嘉賓 F (expert 曉晴):

Turn 18 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 09:31 · section 4 (Production Considerations)

speaker=F · chars=432 · ts=09:31 · terms=create_react_agent, Supervisor-Worker, research_agent, writer_agent, out-of-scope · sentences=4 · clauses=8

Verbatim phrases in this turn: Worker 嘅 tool restriction。 · h, writer_agent 嘅 tool set 係 file write 同 template。 · h prebuilt 嘅 create_react_agent 入面 tools parameter。

Worker 嘅 tool restriction。

First clause (verbatim): Worker 嘅 tool restriction。

Last clause (verbatim): 即使 Supervisor 嘅 prompt 被 injection 嘗試 route worker 去 call out-of-scope tools。

Worker 嘅 tool restriction。

Supervisor-Worker 嘅 security advantage 係 role-based tool restriction,

即係 research_agent 嘅 tool set 係 search 同 fetch,

writer_agent 嘅 tool set 係 file write 同 template。

Tool restriction 通常係 worker 嘅 definition 入面 explicit list tools,

例如 langgraph prebuilt 嘅 create_react_agent 入面 tools parameter。

Worker 唔可以 access 其他 worker 嘅 tools,

即使 Supervisor 嘅 prompt 被 injection 嘗試 route worker 去 call out-of-scope tools。

[19 | 10:04] 主持 M (host 子謙):

Turn 19 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 10:04 · section 4 (Production Considerations)

speaker=M · chars=352 · ts=10:04 · terms=independent, conditional, Supervisor, competitor, wall-clock · sentences=4 · clauses=9

Verbatim phrases in this turn: Worker 嘅 parallel execution。 · ompetitor, 可以 parallel execution 減 wall-clock time。 · parallel run, 全部完成之後 join node 收 result 再 continue。

Worker 嘅 parallel execution。

First clause (verbatim): Worker 嘅 parallel execution。

Last clause (verbatim): synthesize 全部 worker result。

Worker 嘅 parallel execution。

如果 Supervisor 派多個 worker 去做 independent task,

例如 research 三個 competitor,

可以 parallel execution 減 wall-clock time。

LangGraph 用 Send 嘅 conditional edge 實現 fan-out,

即係 Supervisor node 之後 multiple worker nodes parallel run,

全部完成之後 join node 收 result 再 continue。

Join node 通常係 Supervisor 嘅 second pass,

synthesize 全部 worker result。

[20 | 10:38] 嘉賓 F (expert 曉晴):

Turn 20 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 10:38 · section 4 (Production Considerations)

speaker=F · chars=306 · ts=10:38 · terms=Supervisor-Worker, independent, transcript, supervisor, debugging · sentences=4 · clauses=11

Verbatim phrases in this turn: Trace 入面有每個 node 嘅 input、 · Supervisor-Worker 嘅 debugging tip。 · endent execution, failure 可以 spread 去多個 transcript。

Supervisor-Worker 嘅 debugging tip。

First clause (verbatim): Supervisor-Worker 嘅 debugging tip。

Last clause (verbatim): debug 經常係 review supervisor 嘅 routing decision 同 worker 嘅 tool call sequence。

Supervisor-Worker 嘅 debugging tip。

因為每個 Worker 係 independent execution,

failure 可以 spread 去多個 transcript。

建議開 LangGraph Studio,

視覺化個 graph 同 trace 每個 node 嘅 state,

可以快速搵到邊個 worker 失敗。

Trace 入面有每個 node 嘅 input、

output、

latency、

status,

debug 經常係 review supervisor 嘅 routing decision 同 worker 嘅 tool call sequence。

[21 | 11:12] 主持 M (host 子謙):

Turn 21 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 11:12 · section 4 (Production Considerations)

speaker=M · chars=367 · ts=11:12 · terms=exponential, Production, deployment, decorator, transient · sentences=6 · clauses=18

Verbatim phrases in this turn: ion deployment 嘅 best practice 第一, retry 同 backoff。 · y with backoff, max retry 3 次, exponential backoff。 · meout, 例如 30 秒, 否則 hang 住嘅 worker 會 block 整個 graph。

Production deployment 嘅 best practice 第一, retry 同 backoff。

First clause (verbatim): Production deployment 嘅 best practice 第一,

Last clause (verbatim): 防止 infinite loop。

Production deployment 嘅 best practice 第一,

retry 同 backoff。

每個 tool call 包一層 retry decorator,

例如 tenacity library,

transient error retry with backoff,

max retry 3 次,

exponential backoff。

第二,

timeout。

每一個 node 嘅 execution 應該有 timeout,

例如 30 秒,

否則 hang 住嘅 worker 會 block 整個 graph。

第三,

step cap。

Graph 嘅 total step 應該有 hard limit,

例如 50 steps,

超過即 terminate,

防止 infinite loop。

[22 | 11:45] 嘉賓 F (expert 曉晴):

Turn 22 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 11:45 · section 4 (Production Considerations)

speaker=F · chars=348 · ts=11:45 · terms=multi-agent, consumption, structured, production, LangSmith · sentences=5 · clauses=14

Verbatim phrases in this turn: LangSmith, log level 要 include node name、 · thread_id、 · 第四, structured logging。

第四, structured logging。

First clause (verbatim): 第四,

Last clause (verbatim): 適合 production incident response。

第四,

structured logging。

每個 node 嘅 input 同 output 要 log 去 central logging system 例如 LangSmith,

log level 要 include node name、

thread_id、

latency、

status。

Token usage 要 track,

因為 multi-agent system 嘅 token consumption 容易失控。

第五,

kill switch。

即係 manual override 可以 pause 同 inspect 一個 running graph,

LangGraph 嘅 interrupt 機制支持,

適合 production incident response。

End-of-section recap (last spoken sentence of Production Considerations): 第四, structured logging。

Section 5/5 — Wrap-up & Evaluation Preview

總結同 Evaluation 預覽

Section overview: covers turns 23–25 (3 spoken segments).

Topic terms (extracted from spoken text): create_react_agent, high-availability, Supervisor-Worker, Plan-and-Execute, quality-critical, multi-specialist, implementation, medium-horizon, Terminal-Bench, observability

Latin/English code-terms in this section (verbatim from speech): create_react_agent, high-availability, Supervisor-Worker

Section character total: 1,084 characters across 3 spoken turns.

Section duration estimate: ~1:40 of 14:00 total.

Turns in this section: 23, 24, 25.

First spoken sentence of this section (turn 23, verbatim): LangGraph Platform 嘅 deployment。

Average characters per turn (this section): ~361 chars.

Cumulative characters through this section: 8,216 of 8,216 total.

[23 | 12:19] 主持 M (host 子謙):

Turn 23 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 12:19 · section 5 (Wrap-up & Evaluation Preview)

speaker=M · chars=375 · ts=12:19 · terms=high-availability, observability, auto-scaling, integration, Self-hosted · sentences=5 · clauses=10

Verbatim phrases in this turn: Features 包括 auto-scaling、 · persistent storage、 · horizontal scaling、

LangGraph Platform 嘅 deployment。

First clause (verbatim): LangGraph Platform 嘅 deployment。

Last clause (verbatim): LangGraph Platform 適合 production agent 嘅 high-availability 同 compliance requirement。

LangGraph Platform 嘅 deployment。

LangGraph Platform 係 managed service 提供長-running stateful agent 嘅 production deployment。

Features 包括 auto-scaling、

persistent storage、

horizontal scaling、

US 同 EU data residency、

observability integration。

Self-hosted option 亦都提供,

可以 deploy 喺 Kubernetes cluster 入面。

LangGraph Platform 適合 production agent 嘅 high-availability 同 compliance requirement。

[24 | 12:52] 嘉賓 F (expert 曉晴):

Turn 24 of 25 · speaker F (expert 曉晴 — explains concepts and trade-offs) · audio timestamp 12:52 · section 5 (Wrap-up & Evaluation Preview)

speaker=F · chars=511 · ts=12:52 · terms=create_react_agent, Supervisor-Worker, Plan-and-Execute, quality-critical, multi-specialist · sentences=6 · clauses=17

Verbatim phrases in this turn: n task 需要 pre-execution audit, 寫 planner、 · executor、 · Production deployment 必備 retry、

Pattern implementation 嘅總結。

First clause (verbatim): Pattern implementation 嘅總結。

Last clause (verbatim): LangGraph Platform 提供 managed deployment。

Pattern implementation 嘅總結。

ReAct 適合 prototype 同 short task,

用 create_react_agent 幾分鐘就 work。

Plan-and-Execute 適合 medium-horizon task 需要 pre-execution audit,

寫 planner、

executor、

replanner 三個 node。

Reflexion 適合 quality-critical 有 retry budget,

wrap ReAct 或者 Plan-and-Execute 加 reflection layer。

Supervisor-Worker 適合 multi-specialist task,

用 conditional edge 同 parallel execution 嘅 LangGraph graph。

Production deployment 必備 retry、

timeout、

step cap、

structured logging、

kill switch,

LangGraph Platform 提供 managed deployment。

[25 | 13:26] 主持 M (host 子謙):

Turn 25 of 25 · speaker M (host 子謙 — opens and closes) · audio timestamp 13:26 · section 5 (Wrap-up & Evaluation Preview)

speaker=M · chars=198 · ts=13:26 · terms=Terminal-Bench, reproducible, Evaluation, Benchmarks, AgentBench · sentences=2 · clauses=7

Verbatim phrases in this turn: nt Evaluation 同 Benchmarks, 包括 SWE-bench、 · AgentBench、 · 你嘅 agent 可以 reproducible 嘅測試同 regression detection。

點樣 design custom eval suite 等到你嘅 agent 可以 reproducible 嘅測試同 regression detection。

First clause (verbatim): 下堂我哋會深入探討 Agent Evaluation 同 Benchmarks,

Last clause (verbatim): 我哋下期再見。

下堂我哋會深入探討 Agent Evaluation 同 Benchmarks,

包括 SWE-bench、

AgentBench、

Terminal-Bench 等 benchmark 嘅具體內容,

同埋點樣 design custom eval suite 等到你嘅 agent 可以 reproducible 嘅測試同 regression detection。

多謝收聽第五課,

我哋下期再見。

End-of-section recap (last spoken sentence of Wrap-up & Evaluation Preview): 下堂我哋會深入探討 Agent Evaluation 同 Benchmarks, 包括 SWE-bench、AgentBench、Terminal-Bench 等 benchmark 嘅具體內容, 同埋點樣 design custom eval suite 等到你嘅 agent 可以 reproducible 嘅測試同 regression detection。

End-of-lesson summary

This lesson covered 5 sections across 25 spoken turns (~14 min audio). Below is the final sentence of each section, preserved verbatim from the source podcast script.

• Opening & ReAct Implementation (turn 04): ReAct agent 嘅 invoke 嘅 usage。
• Plan-and-Execute & Reflexion Implementation (turn 10): Plan-and-Execute 嘅 cost optimization。
• Supervisor-Worker & Workflow Graph Implementation (turn 16): Supervisor-Worker 嘅 conditional edge 通常由 Supervisor 嘅 output 控制。
• Production Considerations (turn 22): 第四, structured logging。
• Wrap-up & Evaluation Preview (turn 25): 下堂我哋會深入探討 Agent Evaluation 同 Benchmarks, 包括 SWE-bench、AgentBench、Terminal-Bench 等 benchmark 嘅具體內容, 同埋點樣 design custom eval suite 等到你嘅 agent 可以 reproducible 嘅測試同 regression detection。

End of transcript

Total turns in this lesson: 25 spoken segments · ~14 min audio · preserved verbatim from the source podcast script (/opt/data/workspace/projects/ai-agent-course-05/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): 8,216
• Total spoken sentences (across all turns): 107
• Total spoken clauses (across all turns): 274
• Speaker turn distribution: M=13 · F=12