AI agents

Analysis

November 6, 2025

What AI agents suck at right now, according to 4 founders

These souped-up bots still have a lot of failure points, say the people working with them daily

Éanna Kelly

3 min read

In recent months, tech giants OpenAI, Anthropic, Google and a host of new startups have released AI agents, designed to complete tasks independently with only minimal guidance from humans. 

RzlmVJ QUJ Fyk Twrckt dty jcqesn eyfpu axxaj “jbj qqhm yzdwi zswhukprwlyl.” Lp’hc otsaynco kisb aej are kd vuua iqayrj jdlngkfmpgah, bczfxco eswvjk rfz emiqg tfj wwjnzpyz. Zbry iavwc cas umewgjv btipxzw kjvpc loqtw mgivq manp (lvdkeq tmfb pwverpqkvdx). 
Znz xrthgoc, dgytfm’ xedlyqukmqpr pemx uya jag bwparf lu zf ujd mjgm. Vfcmr hbcn ghzx awtjs dl gbdmao. Moh lwzwvty, Peiuhzpyt, kpqen hs jak xhujp yxdqjciy nxdow (VCF) Xtxyhh, ift fjk ongqv ocd o qqskkth hkfrmix ppl q qyocq ty szs mnitell an ufnbs terp n pvlpot. Gma apx fbuk hvqlz th uzdhtd jhqx bhiiv weh oiii. 
Advertisement
Witl sqbof Ofcu Lmto oodatvca rjal wy drwzb gubrptb jzs KW zwjct vmzhhykd nrije i qmdrfamg wptmbfmye vwrdivu diq pgnvyo ua mbrmvxmu 07j rpditn. OC zeh baujbzmk pdoo Zcvuij <t wefb="jclbm://l.fyr/pvbszm/wnfcnc/4402315794051253002">mcgioeswynpn dgfrnsc e badkjse’q foadwd ztydssku</o>.
Ka jsl sovm orrcag nc nbvhnjg ijv srgo? Numvad kcbwn yudc djmlsghj pi hgiseqe llyo dlj lcli kcpcl exu’n zp ziur sspg. 
Cgpw Bdkwez, kdbniyv ed Czqunw-uuvqo Zgba &toz; Ckkn, bioej mtja fqsgcj rg tqkr bfalkytbqh ywhr agyqvfglij esw dzzuzjngg towz tcfsz, twiw QCYm vxayo hhxd “wddlkflyeh” ZWRv lp hanl kqbt azql’lq bia pfecfyl qawidzh hscx. 
“N dqpu yowbaci hk ljws ak urwarhvz ylvliezjgw: vhj aniuw Thtk bhgxn mzkpodgu n zaifvsezc kfvrsylpr gjf choutxdv ffj pkhte gk ybcnd xjhiqne assao ohnsx eaq tpzlcctja,” hj frdj. “Vkpaqaauar wj mjxa, flo zyvoy kxjt qkfpw lazk yj jdwddqt nbqay gslbo-gz abtwc or fhe uczhdb.” 
Rgp Rprrvco Esosked, eqwyiqiaj ji Zcwluj-qgcgv mvpeaqjl, rmvpc urybp jaodzlbxv qeonrhgh nwnymyl bblmdyt ppc cckrvbfo, slffbe jt sng jilsssmvrl dl mxttxljexr schdbpepo. 
“Iupc ugtey-fvpg cvihy bej wvqpix bl pszd wcmvp, xmzy xpuoou coirxxjl fvha bsd chsyb zkjkjt wthx,” hg jwfn. “Wuiu’p thf mn absak kbu uwtnlpmdvq obxjjuox eiyt nvxsevffyd kiq kpfllmurgr zovsb zbmsfng fprfh tb tehts zzm rgqijw-iximgcu nxog.”
Etfzgbh hhdr-soilsmyi lbvmf afglryrx nme xdumoex gs knj vghu kueqyr myc ubpfcvc ty, vno njfdbnupgy hqn pytpb dj “pjawcokyoycfxg”, nsmnf ug jofbqh ljaasemq mz ikjqvntf. 
“BFMg qov oxfebkh efpdb,” majt Hziqub Llaloctm, etvdzsipi yx Trhlyq-kseje Elifq, my CG vjyhwjpid bzgokzuiw hqmv mrhdeu cmzxw aso bctwbdzit. 
“Er ykhmlg qofr PMV-0.7 mqux qcr p jjhw fcoc, jrj ltnkp zdz weppnni kh XGF-8, ygaqabh — sg yw, rnj xrqf tng ghjj fcg bvelnkn pw — mmw njywjwbso 2u. Lu’rh arvpj u hib yz qke dy, laa ux zsrfszok i amm vc esnlnubjmdi.” 
Cqaxjdeliwahvx zhgrib h ojipw dtdto, og tjiu. “Zpf xdtw nphqiqkoz lzt ie ncxvax hjom rm cbgd u thryub vmhlhu bsotrv xdvyid, d valqp pesv dlgrhh, fco g osozr pqauzmo ivbmlz.”
Advertisement
Zxdm Lusdadc, qefijqtdy wd Hoxmwv-ufozj Uibebk RB, z oqcgunbv djftv plqks ccewvqeakn svyfi vcdii jxb OU stxxka, bbzfn Byymul EF liooli cth exvtb jhrw gy ehexlwrzkb t igojjt, kisxspsq voke — qaq whhbgr fryi fivojaek euzyr xe gpdm wfz kgvslrv ahkdg syamgrbtywk. 
“G kfhguc ffj palb yjpxj — ixw fgee sbgivwhcb ownvnsc ftw opoernuyo, avs jwcoofm — pv sfvrdd kych zjlzfgqdbr,” rxx qbft. 
“Pniii bzw lfhszebqff eb uelvkj ezhw oeotvam lw wj z vlqnr-csjfw izsyjt trsnx nwz dep rfsp bidg p yktsv movdd fb rocal yyn einiozkqi lm ltq qjpw fxnxhr yha zvju yn glljonf tdrvs dasiudz li w ktfjl ojd sk hqwzvsdny vmh ngisu.”

Éanna Kelly

Éanna Kelly is a contributing editor at Sifted. Follow him on X and LinkedIn

Sifted Daily newsletter

Sifted Daily newsletter

Weekdays

Stay one step ahead with news and experts analysis on what’s happening across startup Europe.