Guest lecture · COMS W4995 Agentic Engineering
What Gets You Hired, What Gets You Promoted, and What AI Changed
Yan Wang · Superlinear Academy · Ph.D. Columbia '15
QR code for the slides
Follow along
ai-builders.com/resources/w4995
The whole talk answers three questions
01What will get you a job?
02What will get you promoted after you get a job?
03Did AI change the answers?
If you're in CS
If you're switching in
Before AI · every level answers a different question
Junior
Senior
Staff
What people in the role list · AI engineer title, member profilesSkills most distinctive to the title
2024
2025
2026
Machine Learning
LLMs
LangChain
Deep Learning
NLP
RAG
NLP
PyTorch
Source: LinkedIn, Jobs on the Rise (US), editions of Jan 2024, Jan 2025, Jan 2026. Supply side: what title holders list, not what job ads ask for.
What employers ask for · RAG mentions in US AI job postings
2,885 → 12,609
2024 → 2025 · Stanford AI Index 2026 (Lightcast)
How we'll take each one apart
Lasting problem · appreciates
Packaging · depreciates
①Why did it take this shape?
②What changed?
③Split it: ask two questions
If the model were 10x better, would this still matter?
What problem is it actually solving, and do I still have it?
Case 1 · RAG · ① Why this shape
Your documents
01Chunk
02Embed
03Store
04Retrieve
05Stuff into the prompt
Constraint 1 · 2023
Constraint 2 · 2023
Case 1 · RAG · ② What changed
1994
Passage retrieval
2005
Learning to rank
2013
DSSM embeddings
2016
HNSW search
2020
RAG
Case 1 · RAG · ③ Split it
Lasting · appreciates
Packaging · depreciates
RAG · 2020
RAG
chunk · embed · retrieve
If the model were 10x better, would this still matter?
What problem is it actually solving, and do I still have it?
Case 2 · LangChain · ① Why this shape
The layerThe model weakness it patched · 2022–23
OutputParser
AgentExecutor
Chain
Memory
Retriever
Case 2 · LangChain · ② What changed
Absorbed by models, APIs and runtimes
AgentExecutor
Native tool calling · Jun 2023
OutputParser
Structured outputs · Aug 2024
Chain
Models that plan multi-step work
Memory
Long context, plus files the agent reads and writes
What's left: a loop
Modeldecides the next step
Toolruns one action
Resultgoes back into context
What it sees
What it may do
When it stops
When it asks a human
Case 2 · LangChain · ③ Split it
Lasting · appreciates
Packaging · depreciates
LangChain · 2022
LangChain
chains · agents · abstractions
If the model were 10x better, would this still matter?
What problem is it actually solving, and do I still have it?
Case 3 · MCP · ① Why this shape
OpenAI API
Database
Anthropic API
Calendar
Gemini API
File store
Your agent
Issue tracker
MCP
Every vendor format × every tool
One protocol in the middle
The pain · 2024
The design · 2024
Case 3 · MCP · ② What changed
gh pr list --state open
git log --oneline -5
yourtool --help
2,000tokens, down from 150,000
Source: Anthropic Engineering, “Code execution with MCP”, Nov 4, 2025
1980s
Sun RPC
1991
CORBA
2015
gRPC open-sourced
2024
MCP
Case 3 · MCP · ③ Split it
Lasting · appreciates
Packaging · depreciates
MCP · 2024
MCP
servers · clients · JSON-RPC
If the model were 10x better, would this still matter?
What problem is it actually solving, and do I still have it?
Case 4 · Fine-tuning · ① Why this shape · ② What changed
The promise · 2023
Knowledge injection didn't hold up
BloombergGPT · 2023
My knowledge-base experiment
What it's actually for
Job 1 · style that prompts can't pin down
Job 2 · cost at volume
Case 4 · Fine-tuning · ③ Split it
Lasting · appreciates
Packaging · depreciates
Fine-tuning · 2023
Fine-tuning
domain data · weights · LoRA
If the model were 10x better, would this still matter?
What problem is it actually solving, and do I still have it?
And next month's noun is already here
Lasting · appreciates
Packaging · depreciates
Jev · Sept 2026
Jev
options in · probabilities out
QR code for Decision Pac-Man
grapeot.github.io/decision-pacman
Four cases, one pattern
Nobody splits it for you.
RAG
Lasting · appreciates
Packaging · depreciates
a fixed retrieve-then-generate script
The fixed pipeline
Underneath · a course
Information retrieval · linear algebra
LangChain
Lasting · appreciates
Packaging · depreciates
classes that patched weak models
Legacy chain classes
Underneath · a course
Operating systems · feedback control
MCP
Lasting · appreciates
Packaging · depreciates
every tool through the context window
MCP-first
Underneath · a course
Networks · distributed systems
Fine-tuning
Lasting · appreciates
Packaging · depreciates
fine-tuning to add knowledge
Knowledge injection
Underneath · a course
Linear algebra · experimental design
AI-engineer pay premium over non-AI engineers
2024 · 10.7% ENTRY 10.7% 2024 · 15.8% STAFF 15.8% 2024
Levels.fyi, AI Engineer Compensation Trends Q3 2025 · Indeed Hiring Lab, Jul 23 2026 (seniority classified by tasks)
Share of US tech postings, 2019 → 2025
ENTRYMIDSENIOR ≈ −1−7.7+9.0
Percentage points. All US jobs, May 2026: senior postings +14.7% year over year, entry −7.5%.
The strongest objection
“I know it's IR. So I'll build the best RAG.”
Packaging · depreciates
Does a CS degree still matter?
Depreciating
Appreciating
Bachelor of Science
Computer Science
— with seal —
What you memorised
The OSI seven layers, in order
Determinants by cofactor expansion
A red-black tree, by hand
What to keep
Compilers
Principles, Techniques, & Tools
Aho · Lam · Sethi · Ullman
What a CS degree gave you · for after class
Compilers
The intuition
Before AI · my internship
In industry
This semester, with agents
YOUSERVERpay · key A7chargedreceipt losttimeoutretry · same key A7A7 knownsame receipt, no second charge
What a CS degree gave you
Computer networks
The intuition
Before AI · clicking Pay twice
Before AI · autocomplete
This semester, with agents
SHAREDseats left: 1lockTHREAD 1THREAD 2critical sectionwaits for the lockcritical sectionsells the last seat: 1 → 0sees 0, sells nothingwithout the lock, both sell seat 1
What a CS degree gave you · for after class
Operating systems
The intuition
Before AI · the last concert ticket
The same idea elsewhere
This semester, with agents
input size ntimewhat you can afford to wait fornn²2ⁿexact knapsack, worst case
What a CS degree gave you
Algorithms
The intuition
Before AI · a production line
This semester, with agents
ratingsusers × titles≈users × k×k × titlesk = a few dozen hidden tastesLoRA, the same shape: ΔW ≈ B × A
What a CS degree gave you
Linear algebra
The intuition
Before AI · Netflix recommendations
Search
Now · fine-tuning on one GPU
If you're switching in, and if you're not
You already carry intuitions like these
EE · feedback control
The trap
What it gives you
Data science · experiments
The trap
What it gives you
Operations research · optimization
The trap
What it gives you
What the market pays for
IFeasibility
IIDifficulty
IIIDecision
Judgment 1 of 3 · Feasibility
Yes, in half a day
input size ntimewhat you can afford to wait fornn²2ⁿexact knapsack, worst case
No exact solution at scale
Judgment 2 of 3 · Difficulty
Half a day
Decades
Your outline compiler
"Build that in a weekend"
GCC
Judgment 3 of 3 · Decision
CORRECTNESSCAPACITY COSTLATENCY
Knobs that reach it · senior
Choosing the sweet spot · staff
Add a queue
Pessimistic locks
More servers
Student
Solve the problem you were given
Employee
Deliver the task you were assigned
Manager
Decide what to delegate, and check it
01
02
03
Three ways to learn with AI
As a guide
As a sparring partner
As a critic
So, did AI change the answers?
Before AIAfter AI
What gets you hired
What gets you promoted
Depth gets you hired, and the nouns come with it. Judgment gets you promoted. AI made the nouns churn faster and judgment matter sooner.
The engineer
The navigator
  1. 01
  2. 02
  3. 03
  1. 01
  2. 02
  3. 03
On your next build
Before you build
After
QR code for the slides
These slides
ai-builders.com/resources/w4995
Make what lasts.
COMS W4995 · Agentic Engineering · Guest lecture
Yan Wang · October 2, 2026