Slides: Architecting and Testing Controllable Agents: Lance Martin

Source Video

Architecting and Testing Controllable Agents: Lance Martin

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:

aWws

eee)

§ MongoDB o Cloud Ne€O4Y)

OCR text:

WP

ae)

Py a

OCR text:

ED Langcnain

LLM applications follow a contro! flow

Step 1 Step 2

Start —»> 6, —> (() —> et

ae

So Nites Jen vekipeda org win Comrat flo

. nd

| a

, LC

"

OCR text:

C—O

8 e (o ee er emerncrrenterrat CRM 3 wre emma ar aoe 2.8 _ 2). 00 © iia

aS ES . —- en. @

O -—

m Architecting ¢ testing rettoble ogents

. Sohere tngee ong hen

¥ TT \<le en

wh

i aws

, i Tet?

OCR text:

a © —— — ~ Ca

o 2 BS yeree we iersee® ot we “ « 00 © Qi

(is): [from IPython.display import Image, display

class GraphState[TypedDict}:

Represents the state of our graph.

Attributes:

question: question

generation: LLM generation

search: whether to add search

Gocuments: List of documents

wee

question: str

generation: str

search: str

documents: Lististr)

steps: Lististr!

|!

7 f aan ,

i s980Q +: see ww

and | aWws

' ~ eel

OCR text:

WwW OuUargcren

Retrieval is not guaranteed, reasoning harder than retrieval

Ta ae ag sing Cc) Int ate tare inet 8 te stents fats te Sar

vad shal eal teicher athena ee *. or “isaac VO ++ tae ewtee net

- wes, Weeds 0} g

gue weoce @ | OO em teen

} words @ 4 Nereres stare or a

i . heeds #5 58

z | = Si nenve a6 .

5, berets @ §

2 erste @ > ia toe eee

“ e Mreme ae ae

é wma inte nae So nereresessess 300 24800 688972400 EOD 17000 ohne eminem mares

Ups cyouty be UinyyYOGhe«

a A bllps biog langcha'n dev-mult-negdie-in-a-haystacs.

a

- aws

OCR text:

wm OLangcnan

Challenge may be recency bias in LLMs

A likely culprit for this phenomenon is a mismatch between the task LLMs are trained on and context-augmented

generation tasks. Among the documents typically used to pre-train LLMs such as web pages, books, articles and code,

the most informative tokens for predicting a particular token are typically the most recent ones. During pre-training,

this induces a leamed bias to attend to recent tokens. In addition, the rotary positional embedding (RoPE) scheme used

in the open source models we investigate has an inductive bias towards reduced attention at long distances [27] that may

make it even easier for these models to learn to attend preferentially to recent tokens. Extreme recency bias is not a good

prior for context augmented generation tasks where far away tokens may, in fact, contain very relevant information.

ay Reis ak aa a et

|

A aiid an Mi ft @

; gw Microso

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.