InoteveraGaidhi rinotevera
Code Execution Sandboxes yeAgents
Tekinoroji
Nhungamiro yehunyanzvi
Hanisheni yeajenti, kana scaffold, ndiyo kodhi yakatenderedza modhi yemutauro inoishandura kuita mumiririri: inomhanyisa kudana modhi, kuita maturusi ainokumbira, kudyisa mhinduro, uye kusarudza nguva yekumira.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Izvo zvine basa nekuti iyo harness inoumba zvine simba mhedzisiro, saka iyo imwe modhi inogona kukora zvakasiyana zvakanyanya pane imwecheteyo bhenji zvichienderana nemagadzirirwo ayo maturusi, mamiriro uye kukanganisa kubata.
Muenzaniso wemutauro uri wega unoburitsa zvinyorwa. Mumiririri anoda chirongwa chinoramba chichibvunza modhi yacho zvekuita, oita chiito, uye anoratidza muenzaniso zvakaitika. Iyo purogiramu ndiyo tambo. Zvikamu zvaro zvepakati ndiyo loop, tsananguro yezvishandiso, kutumira chishandiso, manejimendi mamiriro, kumisa mitemo uye kukanganisa kubata. Iyo loop inotumira hurukuro kusvika ikozvino kune modhi. Kana iyo modhi ichikumbira chishandiso, harness inosimbisa nharo, inomhanyisa chishandiso, uye inoisa mhedzisiro. Kana iyo modhi ichipa mhinduro yekupedzisira, loop inopera. Mitemo yekumisa inowedzera miganhu: huwandu hwenhanho dzenhanho, chiratidzo kana mutengo bhajeti, nguva yekubuda, kana cheki senge bvunzo dzinopfuura. Tool design ine mhedzisiro yakakura. Iyo SWE-agent purojekiti kubva kuPrinceton vaongorori, yakaburitswa muna 2024, yakapokana kuti vamiririri vanoda mumiriri-kombuta interface yakagadzirirwa modhi, ine mapfupi ekuona mafaera, mirairo yekutsvaga uye nekukasira mhinduro pane dzakaipa edit, uye yakaratidza izvi zvakagadziridzwa mhedzisiro pamusoro pekupa modhi iri goko rakajeka. Mafomati ezvinobuda zvine basa zvakare: kudzoreredza mahombe anobuda, kupfupikisa zvikanganiso zvakajeka uye kudzoreredza zvakarongwa zvinobatsira modhi kuita nemazvo. Context management inosarudza zvinoonekwa nemuenzaniso sezvo nhoroondo inokura: kuchengeta nhanho dzichangoburwa, kupfupikisa ekare, kana kuchengeta manotsi mumafaira. Kukanganisa kubata kunosarudza kana chishandiso chatadza kufona chichikanganisa kumhanya kana kudzoserwa kumuenzaniso seruzivo rwachinogona kuita. Ichi ndicho chikonzero nhamba dzebhenji dzinoda mamiriro. Pamabhenji ekukodhesa akadai seSWE-bhenji, zvibodzwa zvakashumwa zvemuenzaniso mumwe zvinosiyana pane scaffolds, edzazve marongero uye nhanho nhanho. Imwe pfungwa isiriyo yakajairika ndeyekuti nhamba yebhodhi inoyera modhi chete; inoyera modhi uye harness plus marongero. Zvigadzirwa zvakaita sekuraira-mutsara wekodhi maajenti, muchikamu chikuru, akanyatso mainjiniya akatenderedza modhi.
Iyo coding agent harness inopa iyo modhi maturusi ekuona mafaera, kugadzirisa yakatarwa mitsara uye kumhanya bvunzo, zvino zvishwe kusvika bvunzo dzapfuura kana nhanho nhanho yasvikwa.
Mubatsiri wekutsvagisa harness anomisa loop mushure mekufona kwezvishandiso makumi maviri uye anobvunza modhi iyi pfupiso ine masosi, kudzivirira mitengo yekutiza.
Chikwata chinotsiva chishandiso chegomba chine faira mupepeti inoratidza hwindo pfupi remitsara uye inoshuma zvikanganiso zve syntax nekukasira, uye budiriro yemumiriri wavo pamabasa emukati anovandudza pasina kuchinja modhi.
Mutengi anoshanda nemumiririri anobata chero nhare yekudzosera mari iri pamusoro pemari yakatarwa uye anombomira kuti abvumirwe nemunhu asati azviita.
Vanopa mamodheru vanowedzera kutumira zvigadzirwa zvavo zvevamiririri uye makiti ekuvandudza software, ayo anomisa mamwe mapatani ekubatanidza senge maturusi schemas, mvumo uye mamiriro ekugadzirisa. Protocols yekubatanidza maturusi kumamodheru, akadai seModel Context Protocol, inoita kuti kusanganisa kwezvishandiso kuve nyore kushandiswa zvakare. Dhizaini yeHarness ingangoramba iri sosi ine musoro yekusiyana kwekuita, uye ongororo dziri kuenda kunotaura scaffold padivi pemuenzaniso. Kune vavaki, dhizaini yekungwarira, mitemo yekumisa yakajeka uye kutema miti kwakanaka kunoramba kuine basa zvisinei nekuti ndeipi modhi iri pasi.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Hanisheni yeajenti, kana scaffold, ndiyo kodhi yakatenderedza modhi yemutauro inoishandura kuita mumiririri: inomhanyisa kudana modhi, kuita maturusi ainokumbira, kudyisa mhinduro, uye kusarudza nguva yekumira. Izvo zvine basa nekuti iyo harness inoumba zvine simba mhedzisiro, saka iyo imwe modhi inogona kukora zvakasiyana zvakanyanya pane imwecheteyo bhenji zvichienderana nemagadzirirwo ayo maturusi, mamiriro uye kukanganisa kubata.
Iyo harness ndiyo chirongwa chakatenderedza chinoshandura modhi inogadzira zvinyorwa kuita mumiririri.
Mitemo yekumisa inopedza loop zvichienderana nemiganho senge nhanho kuverenga, bhajeti, nguva yekubuda kana budiriro cheki.
SWE-agent akapokana kuti maturusi akagadzirirwa mamodheru anovandudza mibairo pamusoro peganda rakajeka.
Benchmark mamakisi anoyera sisitimu yese, kwete modhi chete.
Kudzosa zvikanganiso zvinogoneka kunoita kuti modhi igadzirise maitiro ayo pane kupedzisa kumhanya.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Code Execution Sandboxes yeAgents
Tekinoroji