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
OCR text:
eee aaa
nov baie
° «
The Agentic Al Engineer
An Al agent is never “done. It lives in a development loop — and the speed of that !oap is the whole
game
i
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
eee aa
MUTAGENT
‘ “ ” 2 *
An agent is never “done.” It lives in a loop.
Seven stages, run over and over — each pass making the agent a littie better. Offline you build & sharpen it; online it
tuns, and reality feeds the next pass.
Conceptualize
<1
cf? ON
{7 Q
Optimize (7) QL. sua
Q) OFFUINE
O 0
Q Q
O one loop Q)
¢) etilice hight cabite tears O
Diagnose Cretuate
Q Q
Q QO
Q 2,
omerat Qa /
CDDOY
Monitor Deploy
7
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
eae mene keen
MUTAGENT
Your coding agent generates the agent.
The signed spec drives a coding agent — Claude Code, Codex, Cursor, Fi, Hermes, whichever you run — that writes the agent
itself. The result is portable: the same agent runs on any harness.
YOUR CODING AGERT AGENT .:
SPEC
4 x © Ga 4
Cmade Code Coden Cursor a hermes
BuURS OW
i
THE SAME AGENT RIJAS O% ARV WARNESS
x @sa a- * Q
® Case PL Hermes, Veroe! Mastre Claude Agente Deepagents
1
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
eee eee ene ee eer
MUTAGENT
Rake $25, 2M
“ 9
Make “good” measurable.
Derive a dataset of cases anda set of criteria — each a binary check, so a fail points straight at the broken dimension. Rolt them
into one 0-100 success rate the loop can chase.
WHA tee Uor-Kere J
EVAL SYSTEM
4 wat 1 PAL ee oe
Tees tea ath the Mee OI ie aida len cence ce aute cueter ys! Hi Hi Hl Hi
TEC teaet pete veatee @ Sa ae aa eR ‘
Lares Fe a Har Ge ata ee Ce MCC ORBOE OL hat Ty 6ee b farhureR WoT IOC r ee Crecnecowce wane
oy
Wo tabricated tects
SAADEH Ska
Correct toce uit
Te tent are He the tee TA eGR arnt ondena tines tetion gies Deen
reams:
bale bey 6
2222222222222
OCR text:
Neeeeeeeee eee eee ann keen
MUTAGENT
FaASE DIAGN IST
Failures cluster into root causes — and become new evals.
Live, the agent is watched. Failing traces are qrauped by root cause, each cause becomes a finding, and each finding is fed back
as a new eval — so the system keeps growing itself.
live production traces Toot causes findings + ree evals
|| missing tool / context eo wen] aseneae ax 2 y seme =
BH Bf ....
wiong eaceation 6 be at one le
a 6 | .
|| || format } scheme oe eae -
wore
EVAL SYSTEN
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
Neeeeeeeeee eee eee ee RR eer
MUTAGENT
Pasi ot DEAGSESE AT STALE
. *
From a flood of failures to a short list of causes.
You can’t read a million failing traces. Diagnosis clusters them by similarity and categorizes each group — so you fix the biqgest
cause first, not the laudest symptom.
SPP TOAS Ce RREOEND CMREES
Ab lm EANRED CALALS
=a ting content es ::
meee ae Cy conten .
IES wrong tool use ee suo
Nn CBBEE ma Off- path / looping ee oe
Seyi
eT) en Sd “0
che BSaesBun Format / polioy muss Pt ze
ata
@ Cluster @ Categorize & Rank
Gaon the facng ter that cot alee atone > ALR Oe Tee eg he Be GHace to orn y > Orde tre cones thy cast - fn thet pgest ona
Tce tit Paar (ead their oe thy cet Deucatt Martti s fleane Calta MO Pent That Ost fom
DO tsacen - coe cluster missing cortert | erong tool - off-path 2 Th caste = top of the queur
f
4 ‘
ete 5
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
Neeeeeeeeee eee EEE
The whole lifecycle, running as an agentic loop.
Define + Design tt
PAL copies ag eat gate ate 2 ,
TapeeN ge eb ad Dt the , é
. mud .
:
Teg Ceca Agent genewtes the agent aie
+ Create the eval eustem = 99% $0.06 é
"238 3 Cao Sfaee . . .
Lene te tata a tte cect tert cae . .
deeded os yd 1 etted from reel sve .
Corrersationa <0" hart oa ttcen we ESP
: Offline optimization loop
Wy abe paler tate Seat tee : :
Deploy
. Monitor
Diagnose > grow
MODE 2 - Live
oRIFT
Fact the re artic
2222222222222
OCR text:
MUTAGENT
: .
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
your machine — while reading from and writing to yout own platforms.
YOUR ERVIRONNENT OE “Sn at
PrSSS? eae tangtuee FP Opentetemetry FY Detedon eB Lanatewin —@ rentrat DE bostor (
Lon, Spay! @ LOCAL = tata never teases
Ww mde cose @ Cote GB icure GR Sy terme
7 Orchestrator
KORO TAT Her gre one MyECT FOr NTEGe
i“ vaqatet
——_______. ————. nite tracet
serel ediagane
: : bediertot os : biageation -
FESEAeIM . sEtEapeM
PROaliw PRiaT oe
raecers Sorte ete ae i vis
were tt erste: Qhemee M Cevselose §@ Cotes Bocuse Gi astra BB deertomrs Claude Verve!
rysgeyyenen POET bed ARCH mONMOE LL Tre kf ite ofehestaton su 6h ence gas
we tat eet + FUTURE CiGJ9 PLATECER
wie gh 2 eG QA Arcoled Uutagest pation o Managed Agents-es-a-Service 10! beaters fay maraget
he ign ment t
“se
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
Cloude Codev2.1.183
Dpus4.81context）
+Clude Kax
/Desktop/Apents/tutagenT/eutagent-
/clear
Read/k/e//
Read 501 Lines
Loaded eutagent-orchestrator/CLAUDE
Ctauding17s161tokes)
(ostext)
41828toke
glbVesion：2.1.183
.tatestversion:
tkropic-ai/claode-cod
irs
Bene
Burak
OCR text:
AOLStage
SYSTE (fron
(ndex)
Nate
Kind
Version
onbcarding
buile
evaluate.
sutagent-ski1t-boilder
mutagent-evaluator
skill
skill
0.1.0-alpha.0
diagnose
autagent-diagnostics
skilt
0.1.0-alpha.8
shared
mutagent-teaplates
skill
orchestrater
eutagent-orchestrator
apest
SETUP/ONBOAROING
complete-googleprovider-workspace mutagent-systes·default
peaini-3.1-flash-Lite-previev-pined_jodge agent-eispatch-obserabitity
claude-code(evalate·diagose)
STATE-
active stage: none
last verdict:none
gated: none pending
*syc.........Indexthe topology ofskilts/agents（scope source foreval/diagose)
Indexxs
Lifecycle
*adit.......Data-flow/leak/context-flo/usdit
(-evaluator]
[-evaluator]
diapose ....RCAcausalchains onrotedfailores-ranked reedies
（5IRPRovE·gatedapply.domstreas of*diagose-ever ato-aplies）
(-diagnostics]
*status.......bere are w-active stagelast verdict·ending gates
StateSetup
help ........Show thisdashbeard
onboard........Unfiedonboarding.config
(atias *config)
Everythings vired andready,Mat ould you Like to shape -evalate，aulit, er *diap
nose?or just tell me in plain wordsvhat yore after,
Bene
Burak
OCR text:
tat
Scriots
evalsate-
nutagent-evaloator
diagrose
mutagent-diagnestics
saitl
skill
0.1.0-alhs.0
0.1.0-algha.8
shared
autagent-teaplates
skitt
orchestrator
sutagent-orchestrator
agest
complete-google providerworkspacemutagent-systea·default
claude-code(evaluate+diapnose)
active stage:none
STATE
tast verdict：none
gated:none ending
Index
Lifecycle
evaloate .....Runealsvaskit/agent-success/faitreverdict (judpe oly)
aodit..........Data-flo/leak/context-flaw/ut aodit
[-evaluator]
[-evaluator]
diapnose.......RCA.causalchains on routed failures-rankedrenedies
（SpRovgated apply.dostrea of*dlagose -ever auto-applies）
{-diagnostics]
StateSetup
*status.......here are v -active stagelast verdict·pending gates
*onbard ......onifiednbearding.config
(alias *config)
Everything°s viret and ready,at vould yoe tike to shape - evaluate, *aotit, er da
isplain wordshat yoe'reafter.
worsec for 35s
11thinking)
Bene
Burak
OCR text:
Term2
Vee
Scrict
QB·5n21.J12:28
STATE
active stage: none
last werdictinone
gated:mone pending
sync........Indexthetopology ofskills/agentsscope source foreval/diagse]
xxx Iodex sos
evaloate.....Ranevalsvsa skill/agest-success/fallere+verdict(judge oely)
mLifecycle
audit.......Data-flow/leak/cetext-flo/uadt
[-evaluator]
diagnose.....CAcausalchains onrouted failores-raskedremedies
[-evaluator]
[-diagnostics]
（SvEgated apply.dostren of*iase-ever ato-aplies）
d515
onboard .......ifiedonboarding.config
(alias *config)
help ..........Shov thisdashboard
Everything's vired and ready.hat woaldyou Like to shape-
Korked for 35s
"diagnose
Routing*diignose-@oLAGNosE（eutagent-diagnostics).
BeforeIsplice atything open,one structuralnote: diagnose normatly consues the faiture bondte the evalater routes te it -but State shous Last verdict:none,There'sno pstrean evalhandover yet.That's
fine: dlagnostics is a mature standalooe stage and can run directly against subject's traces - but Ineedyou te tell me the subject and the source,since I won't guess a target.
-1--（）p-
sioject
Mhat'sthesubject todiaghose-andheredo itstracescomefrom？
1.A systee skill (self-diagnose)
onindeed skills/agnts（skilt-baiteer,
OStICS.
vt
le-g-ctet
iet
Bene
Burak
G3
OCR text:
era ees Scr PC eee
Atlas-Email-Agent
Atlas-Email-Agent
a
a . * a es eet cr a a met
arn ee aa a Se = an ore aa cs ms ane ae eal coo
eS P - - " r 5 ca , : a ss F
o
Coad , eg aed i taheniemaiaalinabans
OCR text:
. ot Pt gan Vis : poo bed a fr fey wma et a er a ee eae
cee : a ete
By
By
in Hea naw ee ee 5
ms Bove cteare ty te ™ ; Bocce Sr ro % rr
ad
eae re J Leal
OCR text:
SDvervier
F-001Th
raft tool is the
F-002Terse conteetreplies are sent
F-004Step-countandcontext tallcomp
Declslons
nensreneaunaaruau
shown below lsa cheaper falbacksgnalNOTaconfldentroot cause,Treat the suspected signalasalead requiringmechanical corroboration
roboration)
Present?
Measure
Declalon
latency-splke
YEX2
output
YES
senin6/040 sm
SA
sn in 4/1946 tmpied
cost-overshoot
YES
snin/1048 sampled
Oerrors
Dfeadback
Scan coverage
LLMDEEP-READ
TRACESNWNDOW
CODE-LEVELSCAN(TIERO)
REPRESENTATIVESAMPLE
(SZAYNY)
100K
n.LLN
p/
102302se
cge.d
（10ofssple)
ineheatmap（colour=avg)
"ace count]
selected:
Bene
oksx120h,
Burak
Copy decielons asmarkdown
Reset all
OCR text:
rea eee eS rere rey co
1,946 pcr y
a
mn uM rv on : ¢ : ” rs rr a“ Oey ” v “ ve tat Ty ”
ary 7 ae
- s - ’ y
iteeg bos ee ae Le eer iv) sf os Dd)
OCR text:
begat darn eonete corey frees ee
1,946 etey2 a s
ES SC Sc 7
an
N
‘¢ oar +
- » aa PSS En area e a = v7 oa we
DR
Cte coe i?
OCR text:
SDvervies
Scan coverage
LLMDEEP-REAS
TRACESIWINDOW
CODE-LEVELSCAN(TIERO)
REPRESENTATIVE SAMPUE
(ANALYZEAS)
s'slicet
LUF
58 ef aindor
mozst/sed/best/and
2021102se
(1ot casgle)
cownge:dl
24htimellne hoatmap[colour-avglatency,number-trace count]
ksat20h9h,07h:latenoyp
Severity
WHAT/WHY/WHERE
Apply
One-liner
latency-spike/tool
tool-defntion
gonerateDraftRepl
-anestedLLMcompositioncaaerazedefter theperent re
-is tho single iwgost coneumarof wallclook
dp-do
system-prompt.
me targetwihnodmotencnergencegudbutlgtiaterevdrnre-dfts
must be preserved.
g-output/r
Onshortudssgetheagentrstotenonile]utputchnn nta
undsrsoec/atem
va ndMesage.andnevr loosupthe conratin-so theconttreciesnothing
cu/coetlayeron toootthedratingboteneck.
Selected:
Bene
Burak
Copy decislans smarkdown
Reset all
OCR text:
Ovevie
F-001Th
De
s40-87%ofend-to-end latencyin28/30traces.
Problem
The agent composes outoour
p.so.the agent effectlvely
decodesthe comersationcontext twicep
[mooel ca],whohare ony2-seach
ng steps
Evldence
traoe673a73b131
Why-chain
Traces arg sow[p5054s,41%oyer 60s]
-taal
salta
Mostof tht timeiginside one tool.genrateDrftReply
.f151
That tool ia a nested LLM call serieRzed attor the pare
I'T-895
Fa-K2 pr-211 S0o
FF-S51
Selected:
Bene
SNOILJN
Burak
Reset all
OCR text:
a A Hl i . E a 7 7 bore eo : g reer:
eY
iN A
soe 7 P|
a y) ' rare
eed Burak . Tet ni aa adeaetneell
OCR text:
Oveview
F-001T
APPLY INSTROCTIONS
1.Set amax_tokens celing on the draft g
2.nable streming so dow
thepipelne alons
mpletesif
Produce
generation
OYTHIS REMEOY
Elminating the neatedg
caching/length tweaksb
ause those onlyshave the
SXON SINL AN
LRY TRtET
Atlas-fmsil-Agent
(clteet
Selected:
Bene
Burak
Copy decislons as markdown
Reset all
OCR text:
Free tey RCE) Pere SES TOy SS
© Decisions — copy approvals
|
%
ae oe 2
7 i &?
BS
ran rs mn $ fl
OCR text:
a
HUTAGENT
' “ #8 a eros Bx
: .
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 cade never leave
your macnine — vehile reading from and writing to your own platforms,
YOUR EMVIRORNENT . eo a
BeEk®> Gt Uaratuee —_P Opentetemetry FY Detedon — B Lanatintn 9 eaintrat «EY Sottor (
Be “gee Oral - Tata never learned
W cite cose @ Cosm EBicure GEM Ay termes
-* Oxchestrator
SOPOT Te Mec go one ayer cer Wage
ferotates
—________ ——— tee traces
evel sdiagnose
: . fasd ater : : Drage ste .
BESEAT IN PESEARIN
PREW Ew PASaIEw
Taggers . RE Ste foe ae
stere tt meites QChomme M Cevde Code B Coder Bcurece GC; tastes EB deertgmrs 1 Claude hy verve
Sanpete Fonte A ROR) mmberfOr sk Lane rr Ofehenteator su sek wane
Ne lms Fwaree PUT ee CiCud PLATEORM
Ate git crete pagent Arosted Utegnet patio — o Managed Agente-me-a-Senice et heaton (a mage
wre tga mart +
y
1aaaaaaaaaaaaaaaaaaaaaaaaamaaaaaaasaaaaaaaaaaaaaamaaaaamasasaaasaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaamaaaamaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaamaaamaaaaamaaaaaamaaaaaaaaaaaaaaaaaaaaaamaaaaaaamaaaamaamaaaaaaaaaamaaaaaaaaaaaamaaaaamaaaaaaaaaaaaaaaaammmmmmama
OCR text:
Meee ee eee ee
MUTAGENT
a”
Mag?
Stop debugging. Start evolving.
Reliable agents come from one thing — a tooo that never stops, run fast enough to compound.
y
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.