Yakha ama-ejenti ocwaningo anothile ngama-Deep Agents kanye ne-Bedrock AgentCore

0 2 7 minutes read

Yakha ama-ejenti ocwaningo anothile ngama-Deep Agents kanye ne-Bedrock AgentCore

Inselele evamile ekugelezeni kokusebenza kocwaningo olunamandla e-AI ukujula uma kuqhathaniswa nomongo. Uma umenzeli wakho efunda amakhasi ewebhu ayishumi, iwindi lengqikithi yalo (inani lombhalo imodeli yolimi olukhulu (LLM) elingacubungula ngesikhathi esisodwa) ligcwala okuqukethwe okungahluziwe. Uma futhi isebenzisa ikhodi yokuhlaziya idatha, ingqondo yesizukulwane seshadi iqhudelana nokucabanga kwamasu kwesikhala esilinganiselwe. Amaqembu ngokuvamile asebenza kulokhu ngokuhlanganisa ama-prompt-chain noma ukucubungula ngokulandelana.

Indlela engcono kakhulu yokudlulisela umsebenzi ojulile kuma-subagents angawodwa abuyisela imiphumela emifushane kuphela. I-LangChain Deep Agents iphatha le orchestration, iveza ama-ephemeral subagents akhethekile futhi ilawula umjikelezo wabo wokuphila. I-Amazon Bedrock AgentCore ihlinzeka ngengqalasizinda edingwa yi-subagegent ngayinye. Lokhu kufaka phakathi isiphequluli sangempela ku-MicroVM (umshini we-virtual ongasindi, onenhloso eyodwa) wocwaningo lwewebhu kanye nendawo ephelele yePython yokuhlaziya idatha. I-AgentCore iyatholakala futhi njengomhlinzeki webhokisi lesihlabathi ku-Deep Agents CLI, ukuze ukwazi ukugijima deepagents --sandbox agentcore ukuzama i-AgentCore CodeInterpreter ngaphandle kokwakha i-ejenti egcwele.

Kulokhu okuthunyelwe, uzokwakha umenzeli wocwaningo oqhudelanayo okhombisa ukuphela kwephethini kuze kube sekugcineni. Lokhu kuhamba kuqondise konjiniyela abakha ukugeleza komsebenzi we-AI enezinyathelo eziningi abadinga izindawo zokusebenzela ezizimele zabasebenzeli babo. Engxenyeni yesi-2 yencwajana yokubhalela, ungakwazi ukusebenzisa le ejenti efanayo ku-Bedrock AgentCore Runtime usebenzisa i-AgentCore CLI, ngakho isebenza njengesevisi ephethwe, ehlukanisiwe neseshini.

Uzokwakha

I-ejenti yakho yomxhumanisi ithola isicelo futhi iqale ihlole i-AgentCore Memory ukuthola imininingwane yocwaningo lwangaphambilini. Ibese iveza ama-subagents amathathu esiphequluli ngokuhambisana ukuze kwenziwe ucwaningo, ngalinye lizulazula kuwebhusayithi yesimbangi ehlukile kweyayo i-AgentCore Browser MicroVM. Lapho ukubuyisela okuthathu okutholakele okuhlelekile, i-subagent yomhlaziyi ithola idatha ehlanganisiwe futhi isebenzisa I-AgentCore Code Interpreter ukuze ikhiqize ishadi lokuqhathanisa nombiko wokumaka. Okokugcina, imininingwane ebalulekile ilondolozwe ku-AgentCore Memory ukuze uthole izikhathi ezizayo. Ungakwazi ukulandelela konke ukuhamba komsebenzi nge-Amazon CloudWatch usebenzisa i-Amazon Bedrock AgentCore Observability noma i-LangSmith.

Uhlobo ngalunye lwe-subagent lufinyelela kuphela isethi yalo ethile yamathuluzi: amathuluzi esiphequluli abacwaningi, amathuluzi otolika womhlaziyi, namathuluzi enkumbulo omxhumanisi.

Umfanekiso 1: Isakhiwo sesixazululo esibonisa ukugeleza kwedatha phakathi kwe-orchestrator ye-LangChain Deep Agents, i-Amazon Bedrock AgentCore Browser MicroVMs, Interpreter, Memory, kanye ne-CloudWatch noma ukulandela ngomkhondo kwe-LangSmith.

Izigaba ezilandelayo zihamba engxenyeni ngayinye isinyathelo ngesinyathelo.

Yakha i-ejenti

Ukuze wakhe le ejenti, ulungiselela imodeli, udale amathuluzi ohlobo ngalunye lwe-subagent, futhi uwaxhume ngentambo nge-LangChain Deep Agents.

Okudingekayo

Ngaphambi kokuthi uqale, qinisekisa ukuthi unokulandelayo:

I-akhawunti ye-AWS enokufinyelela kwe-Amazon Bedrock AgentCore inikwe amandla
Ukuqinisekisa kwe-AWS kulungiselelwe njengokuhlukahluka kwemvelo (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, AWS_REGION) noma ngephrofayela ye-AWS Command Line Interface (AWS CLI). Ukuze uthole izimvume ezidingekayo ze-IAM, bheka umhlahlandlela wokuqala we-AgentCore.
I-Python 3.11 noma kamuva nge-pip noma i-uv yokufakwa kwephakheji
(Ongakukhetha) I-Amazon CloudWatch Transaction Search inikwe amandla ukuze ubuke ukulandelelwa nokubonwa kwe-AgentCore
(Ongakukhetha) I-akhawunti ye-LangSmith yokubonakala
(Ongakukhetha) Isisetshenziswa se-AgentCore Memory esinesu lokukhipha okungenani elilodwa elimisiwe (umthetho onqumayo ukuthi yiluphi ulwazi oluzokhishwa ezenzakalweni zengxoxo)
(Ongakukhetha) Incwadi yokubhalela ye-Jupyter egijimayo endaweni yamasampula e-langchain-aws ukuze kusetshenziswe ngokuphelele

Isinyathelo 1: Setha imodeli

Ikhodi elandelayo ilungiselela i-LLM ehlela umenzeli wakho. Lesi sibonelo sifinyelela kuClaude Sonnet nge-Amazon Bedrock:

from langchain_aws import ChatBedrockConverse
from langchain_aws.tools import create_browser_toolkit, create_code_interpreter_toolkit
from deepagents import create_deep_agent
from botocore.config import Config as BotoConfig
 
model = ChatBedrockConverse(
    model="us.anthropic.claude-sonnet-4-6",
    region_name="us-west-2",
    config=BotoConfig(read_timeout=300),
)

I us. Isiqalo sisebenzisa iphrofayela ye-cross-region inference ukuze uthole ukutholakala okuphezulu. Ungasebenzisa futhi i-ID yemodeli eyisisekelo ngokuqondile.

Isinyathelo sesi-2: Dala amathuluzi esiphequluli

Umuntu ngamunye oqhudelana naye uthola eyakhe i-BrowserToolkit, ehlinzeka nge-AgentCore Browser MicroVM yayo. Uthola ukuhlukaniswa okuphelele phakathi kwabacwaningi abahambisanayo. Ikhithi yamathuluzi iphatha ukuvumelana lapho i-LLM ikhipha amakholi ethuluzi leziphequluli eziningi ngesikhathi esisodwa:

COMPETITORS = [
    ("GitHub", "
    ("GitLab", "
    ("Bitbucket", "
]
 
toolkits_to_cleanup = []
research_subagents = []
 
for company_name, company_url in COMPETITORS:
    browser_toolkit, browser_tools = create_browser_toolkit(region="us-west-2")
    browser_toolkit.session_manager.session_wait_timeout = 60.0
    toolkits_to_cleanup.append(browser_toolkit)
 
    research_subagents.append({
        "name": f"research-{company_name.lower()}",
        "description": f"Researches {company_name} by browsing {company_url}.",
        "system_prompt": RESEARCHER_PROMPT,
        "tools": browser_tools,
    })

I-MicroVM ngayinye isebenzisa isiphequluli sangempela seChromium esixhunywe nge-WebSocket kusetshenziswa i-Playwright (umtapo wolwazi ozenzakalelayo wesiphequluli). Amaseshini ayi-ephemeral futhi ayaphenduka ngemizuzwana. I-session_wait_timeout isethelwe kumasekhondi angu-60 (okuzenzakalelayo: amasekhondi angu-10) ukunikeza ukusebenza kwesiphequluli isikhathi esanele sokuqeda lapho amakholi amathuluzi amaningi asebenza kanyekanye. Amathuluzi esiphequluli afaka navigate_browser, extract_text, click_element, type_text, scroll_page, extract_hyperlinksfuthi wait_for_element.

Umfanekiso 2:Ama-ID amathathu ahlukene eseshini ye-MicroVM aqinisekisa ukuthi i-subagent ngayinye yocwaningo isebenza endaweni yayo ehlukile ye-Amazon Bedrock AgentCore Browser.

Isinyathelo sesi-3: Dala ikhithi yamathuluzi otolika

I-subagent yomhlaziyi isebenzisa i-AgentCore Code Interpreter, i-MicroVM engayodwa esebenzisa indawo egcwele yePython enamapanda, i-matplotlib, kanye ne-numpy efakwe ngaphambili:

ci_toolkit, ci_tools = await create_code_interpreter_toolkit(region="us-west-2")
toolkits_to_cleanup.append(ci_toolkit)
 
analyst_subagent = {
    "name": "data-analyst",
    "description": "Analyzes competitor data, generates charts and reports.",
    "system_prompt": ANALYST_PROMPT,
    "tools": ci_tools,
}

Amathuluzi otolika ahlanganisa execute_code, execute_command, write_files, read_files, list_files, upload_filefuthi install_packages. Udinga imitapo yolwazi eyengeziwe? Sebenzisa i- install_packages ithuluzi lokungeza ngesikhathi sokusebenza.

Isinyathelo sesi-4: Engeza inkumbulo ye-cross-session (uyazikhethela)

Umenzeli womxhumanisi angakwazi ukuqongelela ubuchwepheshe ngokuhamba kwesikhathi ngamathuluzi e-AgentCore Memory asebenzisana ne-API yenkumbulo yesikhathi eside ngokuqondile:

from bedrock_agentcore.memory import MemoryClient
from langchain_core.tools import tool
 
memory_client = MemoryClient(region_name="us-west-2")
 
@tool
def save_research_insights(insights: str, session_id: str = "default") -> str:
    """Save competitive research insights to AgentCore long-term memory."""
    memory_client.create_event(
        memory_id=memory_id, actor_id=actor_id, session_id=session_id,
        messages=[
            (f"Save these research insights:nn{insights}", "USER"),
            ("Insights saved to long-term memory.", "ASSISTANT"),
        ],
    )
    return "Insights saved and are extracted into long-term memory."

Okubalulekile:Insiza yakho ye-AgentCore Memory kufanele okungenani ibe necebo elilodwa lokukhipha elilungisiwe (elifana ne-semanticMemoryStrategy) ukuze ukhumbule isikhathi eside ukuze kusebenze. Ngaphandle kwamasu, i-creative_event igcina imicimbi eluhlaza kodwa ayikho imininingwane ekhishiwe ukuze ibuyiswe.

Lapho umenzeli elondoloza imininingwane, amasu amisiwe e-AgentCore Memory akhipha ngokuzenzakalelayo ulwazi oluhlelekile ngemuva. Ekugijimeni okulandelayo, umenzeli usebenzisa i-recall_past_research ukusesha lolu lwazi olukhishiwe. Ithola amaqiniso abalulekile kanye nokutholwe kwangaphambilini ngaphandle kokucwaninga kabusha kusukela ekuqaleni.

Isinyathelo sesi-5: Dala futhi usebenzise i-ejenti

Hlanganisa izingxenye bese ucela i-ejenti:

agent = create_deep_agent(
    model=model,
    subagents=[*research_subagents, analyst_subagent],
    tools=memory_tools,
    system_prompt=COORDINATOR_PROMPT,
    name="competitive-research-coordinator",
    checkpointer=None,  # Simplification; use AgentCoreMemorySaver for session resumability
    store=InMemoryStore(),  # Internal storage for Deep Agents (separate from AgentCore Memory)
)
 
result = await agent.ainvoke(
    {"messages": [{"role": "user", "content": "Compare pricing for GitHub, GitLab, and Bitbucket"}]},
    config={"configurable": {"thread_id": "session-1", "actor_id": "research-agent"}},
)

I-notebook esetshenziswayo enezinkomba zenqubekelaphambili nesibonisi seseshini iyatholakala ebhukwini lokubhalela elihambisana nalesi sihloko. Isikhathi sokusebenza esilindelwe imizuzu emi-4–6 ngo-Claude Sonnet, okubonisa isikhathi sangempela sokuzulazula kwesiphequluli kuwo wonke amasayithi amathathu. Ukucutshungulwa okulandelanayo kocwaningo olufanayo kungathatha isikhathi esingafinyelela ku-3x.

Landelela futhi ulungise iphutha le-ejenti yakho

I-AgentCore Observability ikunikeza ukubonakala kulesi sakhiwo esinama-ejenti amaningi nge-Amazon CloudWatch. I-AgentCore ikhipha ukulandelelwa nokunwebeka ngefomethi ye-OpenTelemetry (OTEL), ukuze ukwazi ukubuka ukulandelana kwe-orchestration okugcwele ekhasini Lokuqaphela le-CloudWatch GenAI: ukusebenza komxhumanisi ezingeni eliphezulu, ubude bengane ku-subagent ngayinye yocwaningo, kanye ne-subage yomhlaziyi elandelayo. Phakathi nenkathi ngayinye, ungabuyekeza amakholi wamathuluzi nokokufaka kwawo, okuphumayo, isikhathi, kanye nokusetshenziswa kwamathokheni, uqinisekise ukuthi ama-subagents ocwaningo asebenze ngesikhathi esifanayo kusukela kusikhathi sawo sewashi, futhi ukhombe ukuthi iyiphi ikholi engaphansi nethuluzi ehlangabezane nenkinga lapho ukuzulazula kwesiphequluli noma ukuqaliswa kwekhodi kungaphumeleli. Njengokusetha kanyekanye nge-akhawunti ngayinye, unika amandla i-CloudWatch Transaction Search ngaphambi kokuthi ukulandelelwa nokunwebeka kutholakale. Uma usingethe i-ejenti ku-AgentCore Runtime (Ingxenye 2), amathuluzi esikhathi sokusebenza asebenza ngokuzenzakalelayo umenzeli wakho nge-OTEL, ngakho-ke akukho ukucushwa okwengeziwe okudingekayo. Ukuze usebenzise i-ejenti efanayo ngaphandle kwesikhathi sokusebenza, engeza i-AWS Distro ye-OpenTelemetry (ADOT) SDK kanye nelabhulali yezinsimbi ze-LangChain kumenzeli wakho. Ukuze uthole ulwazi olwengeziwe, bheka imibhalo ye-Amazon Bedrock AgentCore Observability.

Ungakwazi futhi ukuthola ikhwalithi yalokhu kulandelelwa nge-Amazon Bedrock AgentCore Evaluations, ehlinzeka ngabahloli abakhelwe ngaphakathi njengezinga lempumelelo yegoli kanye nokunemba kokukhethwa kwamathuluzi. Ukuze uthole imininingwane eyengeziwe, bheka imibhalo ye-AgentCore Evaluations.

Uma uthanda, ungasebenzisa futhi i-LangSmith ukulandelela. Nge-LangSmith, uthola ukulandelela ekugcineni okukusiza ukuthi ulungise lesi sakhiwo sama-ejenti amaningi. Setha okuhlukile kwemvelo okuthathu ukuze uvule ukulandelela okuzenzakalelayo:

export LANGCHAIN_TRACING_V2=true
export LANGCHAIN_API_KEY="your-langsmith-api-key"
export LANGCHAIN_PROJECT="competitive-research-agent"

I-notebook ehambisanayo ibuyekeza le nketho.

Kungani kubalulekile lokhu kwakhiwa

Manje njengoba usubonile ukuthi izingxenye zihlangana kanjani, yingakho le phethini ibalulekile.

Okokuqala, umxhumanisi wakho uhlala egxile ekucabangeni okusezingeni eliphezulu. I-subagent ngayinye yocwaningo iphatha ukuphequlula ngezinyathelo eziningi ngaphakathi futhi ibuyisela isifinyezo esifushane kuphela, igcina umongo womxhumanisi utholakalela ukuhlanganiswa kunokugcwaliswa ngokuqukethwe kwekhasi elingahluziwe.

Okwesibili, uthola ukuhlukana okucacile phakathi kwamakhono. Uhlobo ngalunye lwe-subagent lufinyelela amathuluzi alo kuphela, okunciphisa ithuba lokusetshenziswa kwethuluzi okungahlosiwe futhi kwenza ukulungisa iphutha kuqondiswe kakhulu.

Okwesithathu, ucwaningo lwakho luhamba ngokushesha. Imisebenzi emithathu yocwaningo lwesiphequluli isebenza ngesikhathi esisodwa, ngayinye nge-AgentCore MicroVM yayo, inciphisa isikhathi sewashi lasodongeni uma siqhathaniswa nokucutshungulwa okulandelanayo.

Ekugcineni, i-Amazon Bedrock AgentCore iyimodeli-agnostic kanye nohlaka-agnostic. Amathuluzi e-AgentCore (Isiphequluli, Umhumushi, Inkumbulo) asebenza ngokufanayo kungakhathaliseki ukuthi iyiphi imodeli ewahlelayo. Ungashintsha imodeli ngoshintsho lomugqa owodwa:

# Default: Amazon Bedrock
from langchain_aws import ChatBedrockConverse
model = ChatBedrockConverse(model="us.anthropic.claude-sonnet-4-6", region_name="us-west-2")
 
# Alternative: Anthropic API directly
# from langchain_anthropic import ChatAnthropic
# model = ChatAnthropic(model="claude-sonnet-4-6")
 
# Alternative: Google Gemini
# from langchain_google_genai import ChatGoogleGenerativeAI
# model = ChatGoogleGenerativeAI(model="gemini-2.5-pro")

Sesha umenzeli ku-AgentCore Runtime

I-ejenti owakhe usebenza ebhukwini lamanothi, okusebenza kahle ekuthuthukisweni. Ukuze uyiyise endaweni yokugcina ephethwe ngokuhlukaniswa kweseshini ngayinye kanye nesicelo esizinzile se-ARN, ungayisingatha ku-Amazon Bedrock AgentCore Runtime. I-AgentCore Runtime isingatha umenzeli wakho esitsheni se-ARM64 futhi iqhuba iseshini ngayinye ku-microVM ezinikele amahora angafinyelela kwangu-8. Ngenxa yokuthi i-framework-agnostic, umxhumanisi wakho we-Deep Agents, ama-subagents esiphequluli afanayo, kanye ne-subagent yomhlaziyi konke kusebenza kungashintshile.

I-AgentCore CLI iphatha ukugeleza komsebenzi wokuthunyelwa ngezinyathelo ezine: i-scaffold iphrojekthi nge agentcore createbuyekeza isifanekiso ngekhodi yakho yomenzeli, sebenzisa nge agentcore deploybese ucela nge agentcore invoke. Ngemva kokuthunyelwa, ungasakaza amalogi ngamalogi e-ejenti futhi uhlole ukulandelelwa kwawo agentcore traces. Uma usuqedile, i-agentcore isusa konke okulandelwa i-agentcore deploy idiliza zonke izinsiza ezinikeziwe.

Ingxenye 2 yenothibhukhi ikuhambisa kulesi sinyathelo ngasinye, okufaka izimfuneko kanye nezimvume ze-IAM.

Hlanza

Ukuze ugweme ukuvela kwezindleko, hlanza izinsiza ze-AgentCore ozidalile:

# Clean up browser sessions
for toolkit in browser_toolkits:
    await toolkit.cleanup()

# Clean up interpreter session
await ci_toolkit.cleanup()

Amaseshini esiphequluli aphelelwa yisikhathi ngokuzenzakalelayo ngemva kwehora elingu-1. Izikhathi zotolika ziphelelwa yisikhathi ngokuzenzakalelayo ngemva kwemizuzu engu-15. I-notebook ehambisana nalokhu ifaka ikhodi yokuhlanza esebenza ngokuzenzakalelayo. Uma udale insiza ye-AgentCore Memory futhi ungasayidingi, ungayisusa usebenzisa ikhonsoli ye-Amazon Bedrock AgentCore noma i-API.

Isiphetho nezinyathelo ezilandelayo

Kulokhu okuthunyelwe, wakhe i-ejenti yocwaningo esebenzisa i-LangChain Deep Agents yokucula, i-Amazon Bedrock AgentCore yokwenza okuzenzakalelayo kwesiphequluli, ukutolika kwekhodi, kanye nenkumbulo eqhubekayo. Ukhiphe i-ejenti ku-AgentCore Runtime usebenzisa i-AgentCore CLI ukuze uyiqalise njengesevisi ephethwe ngokuhlukaniswa kweseshini ngayinye, iphoyinti lokugcina elizinzile, nokubonakala okwakhelwe ngaphakathi. Le phethini yokuqoqwa kwedatha efanayo, ukucutshungulwa okuyisipesheli, kanye nokuhlanganiswa kusebenza ekusebenzeni okuningi okungaphezu kocwaningo lokuncintisana:

Ukunakekela ngaphambi kokungenela isivumelwano:Lungiselela ama-subagents ukuze ucwaninge ukufakwa kwezimali, ukukhishwa kwabezindaba, namadokhumenti okulawula enkampani eqondiwe. Isibonelo, shintsha ama-URL esincintisana nawo kumakhasi wokugcwalisa we-SEC EDGAR futhi usebenzise kabusha iphethini ye-subagent yesiphequluli efanayo ngezinguquko ezincane.
Ukudalwa kokuqukethwe:Sebenzisa ama-subagents ocwaningo ukuze uqoqe izinto eziwumthombo ngenkathi i-subagegent ebhalayo ibhala izigaba
I-orchestration yepayipi yedatha:Yenza ama-subagents akhiphe idatha emithonjeni ehlukene, bese udlulisela imiphumela ehlanganisiwe ku-subagent yomhlaziyi ukuze uthole ukujoyina nokuguqulwa

Ukuze uqalise, vula incwadi yokubhalela ehambisana nalesi sihloko bese ulandela indlela yeseli ngeseli. Uma unemibuzo noma impendulo mayelana nalesi sixazululo, shiya amazwana kulokhu okuthunyelwe. Uma unemibuzo noma impendulo mayelana nalesi sixazululo, shiya amazwana kulokhu okuthunyelwe.

Ukuze uthole ulwazi olwengeziwe mayelana nezinsizakalo ezisetshenziswe kulokhu okuthunyelwe, bheka ku: