Markdown source

Slides: The Agentic AI Engineer - Benedikt Sanftl, Mutagent

Source Video

The Agentic AI Engineer - Benedikt Sanftl, Mutagent

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

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
• OCR decision: ready — Diagram slide with multiple small labels around a loop; OCR is a better fit than direct transcription.

Slide text:

HUTAGENT

An agent is never “done." It lives in a loop. Tht loop

Seven stages, run over and over -- each pass making the sgent e littie better. Offline you bulld 8 sharpen it; onllne It runs, and realty feeds the next pass.

Conooptuelre

Cpther

OFplirg

otfllm+ tild - onllae leazn one loop

Monnor ctpioy

TheArentlc Bene Burak 02 1 14

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
• OCR decision: ready — Architecture/process slide with many small labels, icons, and a flowchart; OCR is appropriate.

Slide text:

HUTAGENT

PhaSe 2. BuIlD

Your coding agent generates the agent.

The slgned spec drives a coding agent - Claude Code, Codex, Cursor, Pl, Hermes, whlchever you run -- that writes the agent

Itself. The result Is portable: the same agent runs on any harness.

:SPEC +1: Your cooIuc Acett: AGEnT

Chude Codt. Codx: Cureor:T.:Hrmtt '.

fuas 61

THE SAME AGENT AUAS ON AUY HATHESS

loxal - in yout coding agent clous.manag+d

X F.

Hermes Ytlte Hmtra Ctude Ag+rtt D++phgerits

Th+ At+ntic Bera Burak 6+ 1

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/left-72/opencv-adaptive.
• OCR decision: ready — Dense eval-system slide with multiple panels, labels, and a score table; OCR will be more reliable.

Slide text:

PHASE.3: EVALuATE

Make "good" measurable.

Into one O-ioo success rate the loop can chase. Derlve a datatet of cases and a set of criterla - each a binary check, so a fait points stralght at the bi

Evals + Ihi CeIterIA Lval systln

CUIDED -.COLO STaRT DATASET.:PASS+FAIL PER CASE

Datne crfst'i tth the ttea ftom tre toto asd A fitt touarghtt - ro dete rtedad ye'.

Dertre crlter'a teom peoducton tract tor.: DISCOYEHEO'- FAOH TRAces CAITERIA = EACH A BILARY CHECX Cites lts touro+

Ho tbrcattd faot.

CATaSeTs - THE CaSES Corrtit tooh utd

SyxTntsizr - rnoH. Geouno Ttuth 1 Vetd oe tput tonm!

G+r+rat+ ct++ from t+ te+o helcree tioortet or hnewn-good tamp'h -- b+ter+ you hrrt tetla.

- FRcH TRaces 61/ 00 Jlya ssions

odueten fure irto + regrehntatt

Tho Artntic Bene Burak.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
• OCR decision: ready — Cause-clustering diagram with small labels and counts; OCR is appropriate.

Slide text:

HUTAGENT

asonorio..: t asrhd

Failures cluster into root causes - and become new evals.

as a now oval -- so the systam keeps growing itself. Lve, the agent is watched. Falling traces are grouped by root causo, each cause becomes a findlng, and each flndlng is fed back

live production traces $$snt?. 1001 flrdiags - rt= tvals

msaing tool / contuxt:+ oval: ass+rt tool pr+t+nt

wrorg ttcalatlon +'tval: eicalntion'rule

torrmat / scherm鲁 + tvel: schtea chech'

▼ ptqiosqt

Eval. ststen:

Tha Arntic Bene Burak 11 F 16

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/left-72/opencv-adaptive.
• OCR decision: ready — High-density scale slide with ranked causes, counts, and explanatory boxes; OCR is the right capture path.

Slide text:

PHASE 4: DIAGOSE · AT SCALE

From a flood of failures to a short list of causes.

You can't read a miillon falling traces. Dlagnosis cluaters them by simllarlty and catogorlzes each grou

cause frst, not the loudest symptom.

FAILIRO TAACES A fiu Rauxlo causts

tlust+r: Wrong tool Lihe Otf-path / looping Llulrg conteart

tat+go+iro Format / pohoy mdts Uhupporttd citha

Cluste Categorlre ★ Rank

Croup the tehng trece that look alke hto one buohst = no humus rtdt them one by cmt. Labt cech cutter by thae bnd ot falure, Da trrery buoket mape to a fuxab+e c+uta. Grde thu+ ctut Hrst, rot the lo

+o trac+n + or+ clustot Bissing coatet. arong tool - ofl-path 1.2kc

Tht Artatic Bene Burak

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.97
• Text source: advanced OCR rapidocr-live/left-72/opencv-adaptive.
• OCR decision: ready — Dense diagram slide with multiple small text blocks and labels.

Slide text:

The whole lifecycle, running as an agentic loop. THE PAYOFF ··ALL FIVE PHASES. AS ONE HACHIHE

Dofrmo + Deign. The A Enghisr cigturm mhy erd whet 'otds Goodl' loo'u Pte - Ird turn. it lrto th.

2t Create the evel system The Codlng Agent generaits Ihe sgan? ++tinst the Sp+o. [POOE 1 COLD STANT] 99%. A41+T.$0.05

conltetlont, net hmnd-tuthpntdr. quidt. to 'good'n Btrd Trom ril trom the Seta snd your hatorie taoh Dtrls+ th+ dhtet td toortd crttl ++++rt

rtgrt ton-gated betore it cen dhploy. Fun the delaahrt → Evtuate + Optmir*. Olh+t optlatlon loop Cyey oyait chp a bettt teran

[3] D+ploy Co e lth e feedbtck mtoraritn t pltar.

Thhures tur'eoe se they Teope Montor Lhit imtl res on prod Eodon irlletr

7 ' at now drbiet out the ree wond A rna y:Each felure ie rootc. Degnore + grow the Ie-dehitior ketps t!! HODE2LIYED Otift.

The Artntic Bene Bursk

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
• OCR decision: ready — Architecture slide with dense labels and a boxed system diagram.

Slide text:

HUTAGENT

The systen. wheRe It lrves

It runs in your environment - cloud and local.

Mutagent installs as agents inside the coding agent you already use. The agents run locally - your traces and code never leave *swyoneld umo jnon oi Sutum pue woy, buipee, ol!ym - eulyoew jnon

YOUR EMVIRONHENt you? glatforal · cloud. loca!

+idence 1n SOURCES ca Lrgtut A Openltlemertr B Dtdog G Litnhn B Irehtrrt

r+de lour codl + tocil' - date n+w+t lert+

Chode Code:. o Codm. G cunv E n -O Hrm?

f. oxchrtator

= titct1 tN+rates ，

Cyulgator RESEARCH PREVIE whlp: rost dinpose Dlugrotici RESEARCH: PAEYIE"

2. arott, 1o rolarch, previr

ehere it arit+s TARceTs calda * ca Coaeo Coder G csror codiar stontt -.+d tIoscaDiks [① Kute roptgats cloud-tod

mlt racnd of titoes - mtoeht scene thirn srd lhe erehtstrator rurt tha loop sgtl.

PuNs loCally + Entedt htde taur oodhg tgtmt i A hosted Lutsgert pltterm or Yeragtd Agents-as + Servoe -- tet -hott or fuly mtregtd. then you ment R. future: cloud Platforn

Jh*.Ax+ntic Bene Burak 1% / 14

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/full.
• OCR decision: ready — Dense UI screenshot with small labels, cards, and prompt text.

Slide text:

MUTAGENTAtlas-Emal-Agent-DiagnostioRoport

6547ated2026-06-08122:31:582 souxce langtuse1,946 traces4findings

③.Dvervie F-001Thaaf olist

Atlas-Email-Agent

946trsces·1,508sessions·24h

OAtlas-Emall-Agent

agent

s.aethrdple.s net-4.6-(redacted)-vl:0

t（70.·17.806t clsck

_time>2026-05-06T18:39:10 ems(amexeroising my oming_email>HiAtlas,I would Ike to unsubscribe.Can you also delete allmy Information from routput-the you are done?Thanks,Jordan</incoming_emal>Torespond to the Ionly see messagessent via thesendMessage tool

Selected: Bene 127s Burak 74 1,946 Copy declslons amarkdown Reset all

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.97
• Text source: none.
• OCR decision: ready — Dense UI screenshot with prompt text, metrics, and nested panels.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: none.
• OCR decision: ready — Dense table, cards, and timeline heatmap are better suited to OCR than manual transcription.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: advanced OCR rapidocr-live/full.
• OCR decision: ready — Dense table, cards, and timeline heatmap are better suited to OCR than manual transcription.

Slide text:

Overvien F-001The 00- DDe

Slgnal census→primary selected byfallure mpaotxprevalence

ation]

Signal/fallure-mode Present? Measuro Decislon

latency-spike YES 1790/46[)

sndino-Duseya YES senin5/1046sampled

YES seenin4/g46 smpled

cost-overshoot YES snn3/40 smled

Oerrors

Ofeedback

Scan coverage

5slicesindo-bas 1,946 TRACESNWINDOW DODE-LIVELSCAN(TIERO) 1,946 00-00T REPRESENTATIVE SAMPLE 102 Sof windo p1/1// LLMDEEP-READ covetag:cin 102 (ANALYZIRS) 102-of102sarpled (100%ofsle）

24h timeline heatmap(colouravglatency,number·trace count)

s0d.pr

141

Selected: Bene Severity WHAT/WHY/WHERE Burak One-liner Copy decislons as markdown Reset all

KTGH

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: none.
• OCR decision: ready — Dense table, cards, and timeline heatmap are better suited to OCR than manual transcription.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: none.
• OCR decision: ready — Dense table and heatmap layout are better suited to OCR than manual transcription.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: none.
• OCR decision: ready — Dense product/report UI with many small text blocks and tabs.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: agent_vision.
• OCR decision: ready — Dense screenshot with small text, headings, and a long assumptions block.

Slide text:

Sampling coverage — representativeness proof

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.97
• Text source: none.
• OCR decision: ready — Dense remediation screenshot with multiple small sections and embedded notes.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: agent_vision.
• OCR decision: ready — Dense decisions screen with approval controls, feedback box, and live markdown preview.

Slide text:

Decisions — copy approvals

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.99
• Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
• OCR decision: ready — Architecture diagram with many small labels and product/tool names.

Slide text:

HUTAGENT

The systth. where it lives.

It runs in your environment - cloud and iocal.

Mutagent installs as agents lnside the coding agent you already use. The agents run locaily - your traces and code never leave your machlne - while reading from and writing to your own platforms.

YouR EnvIRonHENT your platforas: clood- local

++id+nce 1n SOURCES Openlelhmetry M Dtdog B lantruat

Clrd+ Coot': o cde. G cunvE 1 a xm ng aont + Lota - datl nwr levte

condcote trw Fteoyde = ont soirt per staoa:Orcbrstrator

i. += titc+1 gtrerates

aroot ebulld C+.l-ator RESEARCH DREYIE +r+l. agent RESEARCH: Diupnostirs sdlagaost: PRETIE!

2 at+att in Itptrth, prtvita

rhere it azit神s TARCltS cod+ [ cavde cooe! o Codax G areor codirg al+nt..ad [ Lata rptgt ] fIs+aoiks coA ro. cloud-Hagd

Targrts run n ntt reurd of treot - mtrhtr cotrs thtm sid th+ ereht'taor rurs the laop *gi'y.

RuNs LOCally': EnBedd! ht'de yeur codrg sgent -- A hcsted Llutagtet pleitora or Lrg+d Agerts ese-fervce -- +eh -haut ge tuhy Turaged. Yh+h you +mt ht.. sUTURE:CLOUD PlaTfORN

Jhe Armtic Bene Burak 1$ /14

• Recreated text/layout view: open HTML recreation
• AI slide classifier: title_card confidence 0.99
• Text source: agent_vision.

Slide text:

Stop debugging. Start evolving.

Reliable agents come from one thing — a loop that never stops, run fast enough to compound.

Hidden Non-Slide Evidence

• `slide-001.jpg` — title_card confidence 0.97; not a content slide
• `slide-009.jpg` — demo_video confidence 0.99; Screen-recorded demo with webcam overlays, not a standalone presentation slide.
• `slide-010.jpg` — demo_video confidence 0.99; Screen-recorded terminal demo with webcam overlays, not a standalone presentation slide.
• `slide-011.jpg` — demo_video confidence 0.99; Screen-recorded terminal demo with webcam overlays, not a standalone presentation slide.
• `slide-012.jpg` — demo_video confidence 0.99; Screen-recorded terminal demo with webcam overlays, not a standalone presentation slide.

Classification audit: raw/sources/slide-ai-classification/slides/pSto5YaNGUo/audit.json

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.