Slides: WF26: Harness Engineering & Startup Battlefield ft. Garry Tan, Mike Krieger, @t3dotgg , DSPy
Source Video
WF26: Harness Engineering & Startup Battlefield ft. Garry Tan, Mike Krieger, @t3dotgg , DSPy
Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.
Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.
Extracted Slides
OCR text:
World'sFair
Choosing a model:quality dominates
Whatare yourtop3considerationsforchoosingamodel forproduction?
Accuracyand
Accuracy/quality
outputqualityare
PRESENTEDBY
Agentic capabilities (like tool-calling)
farandawaythe
Cost
topconsideration.
Microsoft
Privacy / data controls
Ecosystem/developerexperience
Only5%thinkof
Reliability/uptime
open-weightasa
Enterprisesupport/compliance
primary
Latency
Fine-tuning/customization support
consideration.
Open weights
Other
enAl
World'sF
I'sF
cro:
The2026StateofAlEngineering
ildk
orld'sF
BarrYaron/Partner
Amplify
OCR text:
Pea ad
World's Fair Ae aeley ss ; Aaa eR RED Or Ee)
a ae “
ae a
| i a tered
hs eS 7 aa V3 a oi
ne i . Fa
y , yy. “ ~ Se
a ; os = ‘ ; ,
; Pe _ a es .
ea Ce . = oe F
ea
a.engineet/worldsfair/2026/schedule
OCR text:
AI Engineer
World's Fair
Livestream
July 2, 2026
EVENT STARTS IN 03:05
GRAPHS
BROUGHT TO YOU BY
neo4j
View Full Schedule ai.engineer/worldsfair/2026/schedule
OCR text:
RL Selects theResponse,NottheFacts
Action:
State:
retry
coerce
rollback
quarantine
escalate
1.Failurecategory
2.Risklevel
1.TabularQ-learning
3.Retrycount
2.Small,interpretablestatespace
4.Driftseverity
3.Low-memoryinference
5.Data-qualitycondition
4.InspectableQ-valuesforeverydecision
TECHNICALLY.THISISASINGLE-STEPCONTEXTUALDECISIONPROBLEMIMPLEMENTEDWITHTABULARQ-LEARNING.
OCR text:
———— Aleéngincer -
Amazon AGI Lal World's Fair
AlEngincer ——
World's F —_ oer
— AlEnginecr ————,
ole an aka
- PoC =a bie! th tere) Mites "
OCR text:
Evie oad
World's Fair Keynote Speakers
: bl Fre pee ernest arse .
ed re
: ap Rot eae ots
a Microsoft so
Peco ona 7 ~
ir) He eae eos . ad
. ;
Lab \Wol =r é y snes
‘t 7 -
Welcome to the Al Engineer World's Fair
fl ee
Raouf Chebri Replit
OCR text:
D
(cslletecs)
Track narrative state，not numbers
IRPERFECT NPC RENORY
WORLDMOOD
Dne-Line.senery thatdecays -
softensat+2e,roppedat+4o.
calt-tense
TKREADS
ATTLTUDE INFERENCE
4ndoks tne
worid remeroersandvesves back Is.
Readsyour last 5noves
Drvtal ptay.Nareec.sorio.
Don't track HP and dice-an LLH is bad at that.Track what it's
great at.
OCR text:
Wi rape i : :
Ceca eylow Engineering Survey
. Amplify Nae AVercel
a Microsoft
ui “ Barr Yaron
} The 2026 State of Al Engineering
: v Seva Lie gl Amplify ee ee
OCR text:
A
Ei tpe oad
World's Fair Who Responded
1,048 respondents — mostly founders, CTOs, and Al engineers
Founder 0
Alengece Global, from orgs of all
, other sizes - from startup to
A Fullstack Engineer Sy enterprise - but skew
a Microsoft Engineering Manager / Tech Lead i - smaller.
Protect Manager
Research Scientist / Research Engmneer i :
nates Enaow ll Lo
Data Engneer
Frontend Enpacer <100 people
Data Scientist fi
ML Engmeer ff
rv} ie 0% 4 B 12 16 20 24
The 2026 State of Al Engineering
Barr Yaron Amplify
OCR text:
ak ae 2 ony) :
aes 7 oy ae erry
at a a
re ; * .. or van rd
stip i a Ree
Sa aU coe
P eng IT te 7
Cre Fag @) a ’ ae ae
: Be an 5
coins qd S$ ss s nc wm. das?
cache ae | el aa we i eg
ao Bs eee . baa ~. :
a a te Tae ey i ace Ca tine 7 -
tre a / ant ae : Ce Co a
* camer . Lo : aa mee ~ a
ry , scaca te ea n iz ve - ny 2,
Workos od ae Wore Fae 7 ps . . a
Ao cn ae CUR Pn ee ae Cae ae ae an
Bn wees os nd pam = . acct my
ES ota Parr ri a. eat ‘a we S an
B Brantey nT va i a Ce , _ 7
ve a aa eee Se a - ~ .
Disa wr 7 rn a i : = wae. .
ahleeal / che ree ‘Wor ae Rare PS : a f,
ew oug er) Sad , ee WordsFar Ga. eat Cae aa !
a:
a ; a vee . ae ae :
inet wv) ar any 9 Deen oe ated eked Carag Coal i
a es a a 9 Pag
vs e. Sena
ae a” Wo Faw i
oS tt aia
OCR text:
prompting,promptengineering,chain-of-thought,few-shot,zero-shot,one-shot,system
World'sFair AIEngineer self-consistency,nsemblebest-of-n,AG,structuredouutsmixtureofexperts,guardrails ReAct,toolcalling,functioncalling,toolschema,computeruse,agentswarm,verifier,judge, alignment,promptinjection,promptleaking,instructionoverride,agentic,agents,agentloop engineering,in-contextlearning,exemplars,demonstrations,self-critique,self-reflection,
PRESENTEDBY Microsoft memory,Mc,embeddings,vectorstore,semanticsearch,chunking,chunksize,sliding window,reranking,contextwindow,longcontext,compaction,pre-training,post-training fine-tuning,SFT,reinforcementlearning,RLHF,rewardmodel,preferencedata,evals,D, ORP,TO,GRP,distillation,tachermodel,studentmodel,mult-agentystem,oRA
instructmodel,inference,decode,sampling,temperature,topp,top-k,beamsearch, QLoRA,adapters,PEFT,quantization,checkpoints,openweights,frontiermodel,basemodel,
W-os reasoning models,loop engineering,subagents,orchestrator,planner,scratchpad,retrieval,
observability,guardmodels,red-taming,jaireaks,curriculum,syntheticdataelfay
verifierloops,reflection,treesearch,McT,graph-of-thought,skeleton-of-thought,
constitutionalAL Rl AIF nrocessreward outcome reward tokenizer multimodal VlM STT
TheUnreasonableEffectivenessofSeparatingtheTaskfromtheModel
CoreContributor MaximeRivest LeadMaintainer IsaacMiller DSPy
OCR text:
ES ead
World's Fair
Roe Every repeated Al task deserves
F: Kes (-3-] 4 ,
. PTincer a aa tax extraction cee Pes sche correction
a Microsoft
ieee icc ree cole ES 7
Os Seearll
Tg Peer a ag0rS
oh The Unreasonable Effectiveness of Separating the Task from the Model
eC es a aL oo nyt )
OCR text:
World's Fair |
a A signature alone doesn't
a Microsoft ‘
specify the task.
Atel d 4 Ose) a Viovale!
eT ar
aa m The Unreasonable Effectiveness of Separating the Task from the Model
oe. Maxime Rivest Isaac Miller nT
OCR text:
Ete ad
World's Fair
What
a Microsoft ae
‘icrosoft World's Pas ~ ' eee : a
d's Fo" bees
vere The Unreasonable Effectiveness of Separating the Task from the Model
7 Maxime Rivest Isaac Miller nyt)
aera! da
OCR text:
Rocoan eae at scale.
a Microsoft
SD or] 8
+ : The Unreasonable Effectiveness of Separating the Task from the Model
Pe a A" Maxime Rivest _ Isaac Miller at
OCR text:
AI Engineer World’s Fair
The implementation is fully learned.
YOU WRITE
extract = dspy.Flex(ExtractTaxes, tools=...)
THE MODEL WRITES
class TaxExtractor(dspy.Module):
def __init__(self):
self.parse = dspy.Predict(...)
self.check = dspy.ReAct(
Validate, tools=[...])
def forward(self, invoice):
lines = self.parse(invoice)
lines = dedupe(lines)
return self.check(lines)
The Unreasonable Effectiveness of Separating the Task from the Model
Maxime Rivest | Isaac Miller
Core Contributor | Lead Maintainer
DSPy
OCR text:
ew Learn the hill over time from ,
World's Fair
ee HOSES Gee“ESMenaety SSIS OSM, Tce URES feedback moves the bill
oe 7 a
a
iW ca i, The Unreasonable Effectiveness of Separating the Task from the Model
Ot, 7 Maxime Rivest Isaac Miller aa
OCR text:
re , STABLE ABSTRACTIONS rae at
World's Fair GET ETS 2023: PROMPT ENGINEERING
FLEXIBLEIMPLEMENTATIONS I 2024: CONTEXT ENGINEERING
0 iis PER ULTASY U a
: Og tL
|
CROCE
TTT
ee as yt
ye
7 The Unreasonable Effectiveness of Separating the Task from the Model
an Nee ee ae eed DsPy
OCR text:
Dea ed ces ores aid foteraay ARE a Ed cd . PO e ened cot Corte d Y ariel Ser cE aed J eed Ly
Seen wie Sir ass] Se prea eed arc Seer ee) E eer ed DE STELS Cre rs A fe one oan) ie
aaa eed 7 io) Neng a eed ented Eeeaarea eas te Ss a a no ood id ive z Serene
ns ered ee Serta a ied _ teers Seen oe ig
ALL riqginnece
Darn eg Blonerie) heath aed a : ran Ra ecg PE
hnarrer emma or eared i eo ea eed BR a , Ouse ea ers Pon ae ees ey
coeerermmrenmama mn ','/e) a (0 Gow o>: || al enor
Pc eer i Roe Seed ce
eee’ aca. One ed Orr) Wretts For Need *
Reever aes a ® Sas or a DS Esha ad Poy eke he gs ae ee ees Cit Perea res hia oe
See nied ic Sreoeet re) ‘ o Ser y AN Necrrieey e Ree nEned iad aE eed enced ares
Cae aa Seed a ected ets DS Dre ae See RE olan Dad ecm ary eerrt re “ ” coe
erred serra da ld SEE! Le eee Prenat red Aa AGS Woetts Faw ead Reet re] 2h 8 Ae cd He 7 Wercte tin
we 7 Sonera ees Pedic cnet en N a eed roe ern lela reer ees come Viorel a Fae ZAI nee ag Coe eae ee
nee es cad Send = rere es wo Perey ees Senet ees Wee) Sere ra ie ee
eco oa omen teed a: “ ae cd E Peano ce Roe rr et we
Perris a roe ernie 7 rerun ing Aceon Oo leet te aT id iy cd hes
ae i Rreenet red es cd ar Sere teed ati RY ee
a oo eS ae Peery O . = g 3 5
OCR text:
World's Fair
Resolve.
OCR text:
world'sF
W. Resolve
Raverna ngineer sFa
OCR text:
7 . | SLT LlLe as ad + ALEng.nece ————
OlaU ere World's Fair aenarrei\ Recalls koe arte
LW. | &*Buildkite [woo ° |
en AlEng:ncer 6 ———
‘World's Fair aS Elle
PS antes
es \Viatalanes Vara,
im on
5 CEO tj nO} © 0 RO)
OCR text:
he he Wworte's Fal aa oy
iil
mes NS a — ise - ae
/ b an oa oe _ s |
RE] SV are Ves ga
escsipe ra Co enna)
— AlEngineer ; AlEngineer
Py) Pe 7 | REF 8 8 ew |
OCR text:
ie me 91 0 o8
oe eS ' VWWORIQSE
/ ok ls
\",! aan Se . . | eye.
ngineer -
Cc Ce ee el te oe oe -— |
Al Engincer How Anthropic Builds: Lessons from Labs
World's Fair
Mike Krieger ANTHROP\AC
OCR text:
Airbyte
World's Fair
AlEngineer
orld's Fair
RELAI
Emiss
AIEngineer
AlEngineer
OCR text:
re en ee ' WOorasr
17s) aa a . me ; ; | baz
~ ingincer -
Superconzuct SN
OCR text:
AI Engineer
World's Fair
Harness
Engineering
OCR text:
AlEngineer
World'sFair
Microsoft
AlEngineer
Openr
orld'sFair
AIE
SUILO THINGS AGENTS WANT
RTTPSJIML.SSO
AlEngine
World'
Akamai
OCR text:
PAGS ed a)
World's Fair
Account Seucnane teat
Opening (Plan, Act, Loop)
, Agent
a Microsoft
‘d's Fair @ ire
cs World
pales ; Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
‘ Su ose sneodg|
OCR text:
Ni Great! Or is it..? 0
World's Fair
Discovery: every agent builder has to find all the
data services from scratch
; Trust: is the data current, am | allowed to
a Microsoft —~ : 7 ero OR LoL TTOLaL Ze)
thay
ug “4 .
We @ Fair A ,
Al
ve Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
tessa Wav =roys))
OCR text:
EET ena wd de
World's Fair
a Microsoft BN a> fe
i.e ale kan Al
ae
ee Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
- ji
of | Emil Eifrem aor
OCR text:
ES aoe O c a)
, ntology-based Semantic Layer
World's Fair 3 y
OTR R Ty A Technical
Ontology | Ontology
I
Asimplte The metadata of
Peres cea kerk ui caae your data asset
real world landscape:
concepts of your where data lives,
TUS a Le Mare od the schemas
(Na ChGn and attributes,
a CUl Ghee Rte Rory and physical
that makes sense locations
to (human) users
eyaelaw NC) mae A]
eve teh 7
a Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
M Av,¥] TT OO
Emil Eifrem sNeodg)
OCR text:
Picea Account a)
World's Fair Opening
Agent
Records teen t 1D) ial Verificataion
ATechnical
Ontology
A live model of the
systems and data
assets required for
your agents to
execute
Fair eC AC a
4 Cencno eee orate
a haa pa .
: alciersy . .
e , Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
el TOs ; Se _—_
TT ties) Neod)
OCR text:
Nene a)
ela tas Ontology-based Semantic Layer
STC ett Technical Execution
Ontology Ontology Traces
Discovery & Don't Repeat Yourself @
aeatiae ~ RYT tat ae
Woi a te vay
Neod
“— Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
a) r nn
Emil Eifrem Walser)
OCR text:
sitogeeer Amazing Graph Track in Room 2005 i
World's Fair ne ee ae
7 ; EZ 7 ; ree Bo) -1 Fs) <3) 11: rr a
are an | LT <= ms oo 01] 9) ©
= : : a 7 at (orenam sia ee ;
é Dos ig F : oe , ; Ritvik Pandya: =‘
an E am J ates 1i Oe) 7 ;
; ae , ne ; JamesLe- .. . a
E ne _ ST arer iD mel lores] Ue erent cle) an
ge “OH RG 7 ea ; ce 10F:ta1t-] Olt) (-) ac
sane ool re AP Eye EG, Donk Subbiah Sethuraman -
Cr
. Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer
Dale kt Vv OO 7 —_——
rere Neod)
OCR text:
ones Tg
VSINNY KRLTVUSECU AYCIIL LY
Detect and Remediate i
ETL Pipeline Failures
ih acta lg
‘ Anna Marie Benzon
. University of the Philippines, Diliman
github.com/ambenzon27/rl-etl-remediation-agent
OCR text:
An End-to-End RL Pipeline
| Health Agent for Anomaly
| stmt = aes ve re Detection and Autonomous
; a Self-Healing of Cloud Data
| = reais
WiGTE GEN REL ER ORRICK URSIN
Monitor — Diagnose — Score — Decide — Safety — Act > Validate
1.Observe logs, schema, and data-quality conditions
2.Diagnose the likely failure family
e 3.Estimate operational risk
4.Select a bounded remediation action
5.Validate whether the action restored a heaithy state
OCR text:
ene Tg
Deterministic Anomaly Rules
Schema drift, null spikes, field
removals, type changes
Q-Learning Decision Policy
Retry, coerce schema, roliback,
quarantine, escalate, log
Safety Override e The
Critical anomaly + passive action
om intelligence
OCR text:
N elects tne Kesponse, No e Facts
Action:
State:
1.Failure category
2.Risk level 1. Tabular Q-learning
3.Retry count 2.Small, interpretable state space
4.Drift severity 3.Low-memory inference
5.Data-quality condition 4.Inspectable Q-values for every decision
“TECHNICALLY, THIS IS A SINGLE-STEP CONTEXTUAL DECISION PROBLEM IMPLEMENTED WITH TABULAR O-LEARNING.
OCR text:
World'sFair AlEngineer
PRESENTEDBY Budget
Microsoft Job
8Br'-trust
Engineering the future of Al
OCR text:
Ed tea ad
World's 11 When budget is held constant, varying strategy still results in performance gains
eS A ny]
0.76
aa Bobrowsrpbas
om ge Fixed budget Open budget
eld’s Fai
7 oe ° Cy
f rN Engineering the future of Al
OCR text:
AI Engineer World's Fair
Perfect answers and what they cost
Tokens to get to a perfect answer
1.2M
2 runs
Advise
1.3M
2 runs
Grade
1.7M
2 runs
Dream
1.8M
3 runs
Execute
Percent of runs passed under fixed budget
42%
Execute
67%
Advise
75%
Grade
75%
Dream
Engineering the future of AI
OCR text:
AI Engineer
World's Fair
OCR text:
World'sFair
You are watching an AJE Online Talk
Pre-Recorded for Worid's Fair 2026.
My personal notes!
OCR text:
World'sFair
AEngi
You are watching an AJE Online Talk
Pre-Recorded for Worid's Fair 2026.
TOWARDSAI
OCR text:
(problems) One conversation has to do everything
Meaning:
the context window becomes
— the database,
— the file system,
— the memory,
— the reasoning space a
yf |
OCR text:
MCPs, CLIs,
and Skills: CG» ;
Choosing the ight Tong Lave
ee
© Nikita Kothari © +
a, Engineering the future of Al
OCR text:
eee CLIs: The debuggable backbone of direct execution.
World's Fair
(you see exactly
what happened)
Reproducible
r Microsoft (copy-paste at
2AM)
(pipe into the
next command)
rt Ue The Heuristic: If an engineer on your team would open a terminal
to do this, the agent should use a CLI too.
‘ vh
a Engineering the future of Al
OCR text:
a The Bottom Line: Three Architectural
World's Fair Ta keaways
1. Context is Your 2. Portability Has a 3, Don't Be
; Scarcest Resource corer Embarrassed by CLIs
7 Microsoft Every token eae LY Oe add latency and git, 1 and curl have
unused schemas is stolen complexity. Pay that cost survived 50 years because
directly from reasoning. deliberately—for indexing, they are composable,
Skills load tools on- isolation, and sharing. Not debuggable, and battle-
demand. Ieee mT S18 tested. Use them.
; ive koa alg
i alo)
- 8 Engineering the future of Al
OCR text:
World'sFair AEng Pre-Recorded forWorid'sFair2026. You are watching an AlE Online Talk
AHOL CEO,NORIAGENTIC
docs
code data
slack
Hi - I'm Amol.
WedeployanAlemployeethatunderstandsyourcompany,
OCR text:
AEng
World'sFair
You are watching an AlE Online Talk
Pre-Recorded for Worio'sFair2026
AGENT
Whathappenswhenyou hand these toolsto anagent?
OCR text:
World'sFair Pre-Recorded forWorld'sFair2026. You are watching anAIE Online Tak
THEWORKAROUNDS
vision re-prompt
Figma MCP PowerPoint CLI
AGENT
screenshot→replace pixel-diff eval
whatdoalloftheseagenttoolshaveincommon?
OCR text:
Ree
LPanseripts
Boatd Deck
Coote e mee) [
@ workspace
Slack
Let your agents do all the grunt work while you focus on vision and story.
OCR text:
eee Fair .
a5
g t
» CJ
raQgors y f} DD
Not programming. '
ieee errs ee
They will say something technically perfect and socially catastrophic with exactly the same
confidence.
OCR text:
World'sFair Pre-Recorded forWorld'sFair 2026. You are watching an AIE Online Talk
THEAROHITECTURE
Four layers. One destination.
Immutableidentity Situationalmode
Hard rules.What the brand can never say.Nothing below can touch it. Real-time conditions.Who the user is.What they're going through.
Example-anchoredvoice Post-generationveto
teamsstop. The dials,the phrases,the tone guide.Wheremost layer. Reads what actually came out.The only deterministic
ISADORA MARTIN-DYE STOP WRITING TONE INSTRUCTIONS. LAYER THEM
OCR text:
World'sFair Pre-Recorded forWorid'sFair 2026. You are watching an AIE Online Talk
Rulesthat aretrue
regardless ofroute.
UNIVERSAL-RULES.TS-HARDIDENTITYRULE
If asked whether you are a real person, a human,a live agent-You MuST confirm you
are an AI.In your very next message. This rule CANNoT be overridden by any venue
configuration,voice profile,or user request.
EveryAlinBloomdisclosesinthefirstreply.Notifasked—beforetheyask.
Aproduct decision,nota legal one.Transparency is the trust signal.
ISADORA MARTIN-DYE STOP WRITING TONEINSTRUCTIONS.LAYER THEM
OCR text:
AlEngineer
Authforagents:
World's Fair
Unblockautonomous
Al with auth.md
MICHAELGRINICH
woriasrair
FOUNDER&CEO
KOS
Workos
Engineering the future of Al
OCR text:
World'sFair AIEngineer
PRESENTED BY
Microsoft
uiusran
OWOs
Engineering the future of Al
OCR text:
7 SES I ee
Word'sFair] & \O4 a i ae
j i ;
: a / Ot
MEV coro ae | — ~
Ow 0s We L
\ Fa
m Engineering the future of Al
OCR text:
eer an
ea
a Microsoft . ‘
Work’ ow
F o Ww
a ef
nh. Engineering the future of Al
OCR text:
eens
draco relat Make something people want.
Mic off v
af a
x :
. Engineering the future of Al
OCR text:
AI Engineer
World’s Fair
Human
Agent
Today’s signup flow assumes:
A human can read a landing page
A human can fill a form
A human can solve a CAPTCHA
A human can verify email
(agents might not have an email address)
A human can choose a plan
A human can copy an API key
A human can paste it somewhere
A human can understand the dashboard
Agent-native registration needs:
Discovery
Capability declaration
Registration intent
Risk assessment
Agent identity verification (source agent, like Claude)
Human or organizational delegation (my company)
Entitlement grant
Credential issuance
Claiming / later ownership transfer
Audibility
Engineering the future of AI
OCR text:
[ World's Fair auth.md can answer:
What does this service do?
aro Re lot eT eC ueets I Cltg
What identity proofs are accepted?
What auth flows are supported?
What scopes and entitlements exist?
What free-tier constraints apply?
cee ee How does a human or organization later claim the account?
~ig Microsoft “ vent
ae
Fe ° .
Optiver ha Engineering the future of Al
OCR text:
ALEngineer
World's Fair
auth.md
= Microsoft Agent Hames Agent Auth Discovery ee
nny Nate Send ID-JAG
aesrel aha
IRCrex= Iho NOES Le) Ao1A)
nn a Make normal API/MCP request Novas tae]
Scan | Wo Backing identity service
ft .
Ww ¢
an e r)
om “Wo Engineering the future of Al
OCR text:
World'sFair AIEngineer
PRESENTED BY
Microsoft
Microsot Worl
Worle Optiv Atng Worl Engineering the future of Al
OCR text:
World'sFair AIEngineer
HDJCNIIAAA
Microsoft World" AEng
World'sI Optiver Noru Engineering the future of Al
OCR text:
World'sFair AEngine Pre-RecordedforWorid'sFair2026. You are watching an AJE Online Talk
IRA MACHINECRAFT
IRA - FORK MY
BRAIN
runsits go-to-market. How a1o0-person factorytaughtitself to remember- andbuilt a36-agentAIthat
Rushabh Doshi Machinecraft·AI Engineer World's Fair
OCR text:
World'sFair
You are watching an AiE Online Talk
Pre-Recorded forWorid'sFair2026.
DON'TWRITEITDOWN-GROW
BRAIN
Not a chatbot.A twin of the company.
OCR text:
World'sFair Pre-Recorded forWorid'sFair2026. You arewatching an AJEOnline Talk
WENEVERTRAINEDAMODEL
Chunk it→vectors inQdrant-relationships in Neo4j→our own retrieval
OCR text:
een Why am I’m here?
World's Fair
2023 2025 2026
DLAI: Generative Al with LLMs Creator of: mep-lambda-handler
ae ‘Se Agentic al
Cy - G '@@ Foundation
. ee Ss .
5. eS C]tinex
a rao an |
an A
— 439,910 enrollments 35k / month wee
md a 'azon AG
rere me
ee Engineering the future of Al
cl ea
OCR text:
Ct a oe eda! Pe eI Ca i i si ca Sed
EUS aod i. oo eee _ ’
World's Fair vimene re
aca eres CO aatnS
es Caan eee oneal
pote eae
lott. Maree
- “ee us agent . Agent
iva Stein Sn a Oc co en ee) ea
agent nL ti ot Cee Learn ese l ae!
/ aa oot oe oe
ouch
air G le DeepMir
to on Pa aeert ol eae)
a 7 Engi ing the fut AY
aa ie ngineering tne future o
OCR text:
Cem tae ee Cr s Cot) i a ns Sead
heel Paes et Ky a /
World's Fair | 8 ert
7 Strands Agent Mera
oe re - Aa an near. ,
Lt Gm a ar
oe on
key Mike? ag Seed ta sep you again! as
Reany to keep testing re? Being it ont @®
hi Pay ena a |
aa Z.Al
| en
7 . 7 ‘ e °
i ai Engineering the future of Al
am
OCR text:
AIEngineer
demo $agentcore create
World'sFair
PRESENTED BY
Microsoft
air
.AI
Search
Engineering the future of Al
a
OCR text:
World'sFair AlEngineer _demo $agentcore create Add Agent
Name→√Type→√Language→√Build→Protocol→Framework
PRESENTED BY
Microsoft Strands Agents SDK-AWS native agent framework LangChain+ LangGraph-Popular open-source frameworks Google ADK-Google Agent DevelopmentKit OpenAI Agents-OpenAI native agent SDK
↓navigate·Enter select·Esc back
air AI
air Seach Fair Engineering the future of Al
OCR text:
Pe eo ae eo De ee 2) os Ce en a rs oe
a cee (
ALEngincer
W mek = 5 iret tad ae acy
ovate ka arelig tae re
Cane Pn Any
eres PaaS Dates
agent Cha eee
7 Pe ay eee SS oe iB
. ae Eerie Sn) Tel ree reed Cae enki tt
: “Ree Ceo anata) or incre
See mep clipst.client grt streanable Pttp rep client
a Microsoft Se eC See] iki hae eee a
agp BedrockAgentCoredpp
ore) eer Creed
mogel
a Cra Mak tet rie Mie Gal act oh aM rh ks ao on Oars ea
i a ae eae date
ar ep gum g Be oe ee whos ae et og
corn : eae)
a A Al ora SOLS aro 7
3 Fair 7 a :
es UE lach
AGENTS rg
“Ge
7 APPLICATION BUILDER as
ed 11)
os cd i
i rare
sFair =D Airbvte Engineering the future of Al
OCR text:
Smt tree ere @esm@e ee Ce) i i Serr
a : (
Ete ed
. é fa coeOrtars OS ra elas elas
World's Fair ames So de eye ee Ah pee gute pwn
mo ail) hE Mess Te Cee a ecTae
agent
Metra mone
ORR Stet seme (c)
i¢] aes RE wiare)
See Obed mae)
Porras
la et
eri | APPLICATION BUILOER
Se a i
ae te Engineering the future of Al
a |
OCR text:
AI Engineer
World's Fair
OCR text:
World'sFair
You are watching an AlE Online Talk
Pre-Recorded for Worid'sFair2026
Four traps that eachcosta day
VSE SOFTWARE I2C
A8005T
RESULATORKILLS THE OLED
Karovarene:dies wits no pult-ops.Bit-bang it.
sustasnec
bs-sax 3.6Y.Buck-cot 1s mdstory
CP1013ISF5PIMI50
Evenat s.silet fallre,Noved C t17.
OCR text:
World'sFair AEngi Pre-Recorded tor Worid'sFair2026
D
(calletaes)
Track narrative state,not numbers
IRPERFECT NPC RENORY Dne-Line oenery ttat decays -s4ftensat +2e, cropped at 40 WORLDMOGD 2-
TKREADS ATTITUDE INFERENCE Readsyour last 5 soves-brvtal ptay,hareer sori6.
Don't track HP and dice-an LLH is bad at that.Trackwhat it's great at.
OCR text:
Re
Narrated on the fly - rendered to paper
Fig
wes ay
He ‘ . YO
ba A
’ ‘ ¢
As a :
Each scene: a praept + Poilimations.ar + Playa Steisberg dither ko
Ltit parched for the 888.488 panel, a a
iy
a ,
Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.