1 
2AI-agenty pereshli ot issledovatelskikh demo k proizvodstvennym sistemam. Ozhidaetsya, chto bolee 60% korporativnykh AI-prilozhenii budut vklyuchat agentnye komponenty k 2026 godu. No sozdanie agentov s nulya  -  upravlenie tsiklami instrumentov, sostoyaniem, pamyatyu, obrabotka oshibok i koordinatsiya neskolkikh agentov  -  eto slozhnaya zadacha. Imenno dlya etogo sushchestvuyut freimvorki.
3 
4V 2026 godu dominiruyut chetyre freimvorka: **LangGraph**, **CrewAI**, **OpenAI Agents SDK** i **Claude Agent SDK**. Kazhdyi iz nikh prinimaet printsipialino drugoi podkhod k odnoi i toi zhe probleme: dat LLM sposobnost rassuzhdatt, planirovat, ispolzovat instrumenty i sotrudnichat.
5 
6## Obzor
7 
8| Aspekt | LangGraph | CrewAI | OpenAI Agents SDK | Claude Agent SDK |
9|--------|-----------|--------|-------------------|-----------------|
10| **Razrabotchik** | LangChain | CrewAI Inc. | OpenAI | Anthropic |
11| **Arkhitektura** | Na osnove grafov | Na osnove rolei | Na osnove peredach | Avtonomnyi tsikl |
12| **Filosofiya** | Maksimalnyi kontrol | Komandnoe sotrudnichestvo | Minimalnaya abstraktsiya | Daite agentu kompyuter |
13| **Yazyki** | Python, TypeScript | Python | Python | Python, TypeScript |
14| **Podderzhka modelei** | Lyubye (OpenAI, Claude, lokalnye) | Lyubye | Lyubye (nesmotrya na nazvanie) | Tolko Claude |
15| **Zvyozdy na GitHub** | ~29k | ~40k | ~21k | ~6k |
16| **Luchshe vsego dlya** | Slozhnykh workflow s sostoyaniyem | Spetsializatsii neskolkikh agentov | Routinga i triazha | Kodirovaniya i zadach s failami |
17 
18## LangGraph: Stroitel grafov
19 
20LangGraph modeliruet workflow agentov kak **orientirovannye tsiklicheskie grafy**. Vy opredelyaete uzly (funktsii, vypolnyayushchie rabotu) i ryobra (perekhody mezhdu nimi, optsionalno uslovnye). Sostoyanie protekaet cherez graf i sokhranitsya cherez checkpointing.
21 
22Eto naibolee yavnyi i kontroliruemyi freimvork  -  vy sami podklyuchaete kazhdyi shag.
23 
24```mermaid
25graph LR
26    Start --> Router[Router Node]
27    Router -->|needs research| Research[Research Node]
28    Router -->|needs code| Code[Code Node]
29    Research --> Synthesize[Synthesize Node]
30    Code --> Synthesize
31    Synthesize --> End
32```
33 
34### Klyuchevye kontseptsii
35 
36- **StateGraph**: opredelenie grafa s tipizirovannnym sostoyaniem
37- **Nodes**: funktsii Python, preobrazuyushchie sostoyanie
38- **Edges**: svyazi mezhdu uzlami, mogut byt uslovnymi
39- **Checkpointing**: vstroennaya postoyannost dlya dlitelnykh workflow
40 
41### Primer koda
42 
43```python
44from langgraph.graph import StateGraph, MessagesState, START, END
45from langchain_openai import ChatOpenAI
46 
47llm = ChatOpenAI(model="gpt-4o")
48 
49def call_agent(state: MessagesState):
50    response = llm.invoke(state["messages"])
51    return {"messages": [response]}
52 
53def should_continue(state: MessagesState):
54    last = state["messages"][-1]
55    if last.tool_calls:
56        return "tools"
57    return END
58 
59def call_tools(state: MessagesState):
60    # Execute tool calls and return results
61    results = []
62    for tool_call in state["messages"][-1].tool_calls:
63        result = execute_tool(tool_call)
64        results.append(result)
65    return {"messages": results}
66 
67graph = StateGraph(MessagesState)
68graph.add_node("agent", call_agent)
69graph.add_node("tools", call_tools)
70graph.add_edge(START, "agent")
71graph.add_conditional_edges("agent", should_continue, {"tools": "tools", END: END})
72graph.add_edge("tools", "agent")
73 
74app = graph.compile()
75result = app.invoke({"messages": [{"role": "user", "content": "What's the weather?"}]})
76```
77 
78### Silnye storony
79 
80- Tonkii kontrol nad kazhdym shagom i perekhodom
81- Vstroennyi checkpointing i human-in-the-loop
82- Polnyi paritet s TypeScript
83- Rabotaet s lyubym provaidderom LLM
84- Luchshe vsego podkhodit dlya slozhnykh workflow s uslovnym vetvleniem i tsiklami
85 
86### Slabye storony
87 
88- Krutaya krivaya obucheniya  -  nuzhno ponimat konteseptsii teorii grafov
89- Mnogo koda dlya prostykh sluchaev  -  bazovyi agent trebuet bolshe boilerplate, chem drugie freimvorki
90- Otladka potokov grafov mozhet byt slozhnoi bez LangSmith
91 
92### Tseny
93 
94Open-source (MIT). LangSmith (upravlyaemaya platforma nablyudaemosti) imeet platnye tarify dlya proizvodstvennogo monitoringa.
95 
96## CrewAI: Sobiratel komand
97 
98CrewAI ispolzuet chelovecheskuyu metaforu: vy sobiraaete **komandu** spetsializirovannykh agentov, kazhdyi s **rolyu**, **tselyu** i **predystoriyei**. Agenty sotrudnichayut nad **zadachami** s pomoshchyu **instrumentov**, koordiniruemye **protsessom** (posledovatelnym, ierarkhicheskim ili konsensusnym).
99 
100Dumaite ob etom kak o naime komandy, gde u kazhdogo chlena est opredelyonnaya dolzhnost i spetsializatsiya.
101 
102```mermaid
103graph TD
104    Crew[Crew Manager] --> R[Researcher\nRole: Find data\nTools: WebSearch]
105    Crew --> W[Writer\nRole: Write content\nTools: FileWrite]
106    Crew --> E[Editor\nRole: Review quality\nTools: FileRead]
107    R --> Task1[Research Task]
108    W --> Task2[Writing Task]
109    E --> Task3[Review Task]
110    Task1 --> Task2 --> Task3
111```
112 
113### Klyuchevye kontseptsii
114 
115- **Agent**: persona s rolyu, tselyu, predystoriyei i instrumentami
116- **Task**: zadanie s opisaniem, ozhidaemym rezultatom i naznachennym agentom
117- **Crew**: gruppa agentov, rabotayushchikh vmeste
118- **Process**: strategiya vypolneniya (posledovatelnaya, ierarkhicheskaya, konsensusnaya)
119- **Flow**: uroven orkestratsii na osnove sobytii dlya svyazyvaniya neskolkikh komand
120 
121### Primer koda
122 
123```python
124from crewai import Agent, Task, Crew, Process
125 
126researcher = Agent(
127    role="Senior Research Analyst",
128    goal="Find comprehensive data about the given topic",
129    backstory="You have 10 years of experience in technology research. "
130              "You are thorough and always verify facts from multiple sources.",
131    tools=[web_search_tool],
132    verbose=True,
133)
134 
135writer = Agent(
136    role="Technical Writer",
137    goal="Create clear, engaging technical content",
138    backstory="You write for a developer audience. "
139              "Your articles are practical and include code examples.",
140    tools=[file_tool],
141    verbose=True,
142)
143 
144research_task = Task(
145    description="Research the latest developments in WebAssembly in 2026. "
146                "Focus on WASI, Component Model, and production use cases.",
147    expected_output="A structured research document with key findings and sources.",
148    agent=researcher,
149)
150 
151writing_task = Task(
152    description="Write a blog post based on the research. "
153                "Include code examples and Mermaid diagrams.",
154    expected_output="A complete blog post in Markdown format.",
155    agent=writer,
156    context=[research_task],  # Writer receives researcher's output
157)
158 
159crew = Crew(
160    agents=[researcher, writer],
161    tasks=[research_task, writing_task],
162    process=Process.sequential,
163    verbose=True,
164)
165 
166result = crew.kickoff()
167print(result.raw)
168```
169 
170### Silnye storony
171 
172- Intuitivnaya abstraktsiya na osnove rolei  -  legko ponyat
173- Bolee 100 vstroennykh integratsii instrumentov
174- Obshchaya pamyat mezhdu agentami (kratkosrochnaya, dolgosrochnaya, sushchnostei)
175- Samoe bolshoe soobshchestvo (~40k zvyozd na GitHub)
176- Ierarkhicheskii protsess s agentom-"menedzherom", kotoryi delegiruet i validdiruet
177 
178### Slabye storony
179 
180- Menee tonkii kontrol, chem u LangGraph  -  vy opredelyaete roli, a ne tochnye puti vypolneniya
181- Ierarkhicheskii protsess mozhet byt nepredskazuemym, kogda agenty ne soglasny
182- Otladka razgovorov neskolkikh agentov slozhnee, chem potokov odnogo agenta
183 
184### Tseny
185 
186Open-source yadro (besplatno). CrewAI Platform: 99 $/mesyats (Teams) do 120 tys. $/god (Enterprise). Tsena zavisit ot aktivnykh komand i ezhemesyachnykh vypolnenii.
187 
188## OpenAI Agents SDK: Router
189 
190OpenAI Agents SDK (dukhovnyi naslednik Swarm) fokusirruetsya na **peredachakh**  -  agenty peredayut razgovory drugim spetsializirovannym agentam. Eto samyi minimalisticheskii freimvork: agenty, instrumenty, peredachi i guardrails. Vot i vsyo.
191 
192```mermaid
193graph LR
194    User --> Triage[Triage Agent]
195    Triage -->|billing question| Billing[Billing Agent]
196    Triage -->|refund request| Refund[Refund Agent]
197    Triage -->|technical issue| Support[Support Agent]
198    Billing --> Response[Response]
199    Refund --> Response
200    Support --> Response
201```
202 
203### Klyuchevye kontseptsii
204 
205- **Agent**: model + instruktsii + instrumenty + peredachi
206- **Handoff**: peredacha drugomu agentu (modeliruetsya kak instrument, kotoryi LLM mozhet vyzvat)
207- **Guardrail**: validatsiya vkhoda/vykhoda, zapuskaemaya parallel no s agentom
208- **Runner**: vypolnyaet tsikl agenta
209- **Tracing**: vstroennaya nablyudaemost dlya vsekh vyzovov LLM, ispolzovaniya instrumentov i peredach
210 
211### Primer koda
212 
213```python
214from agents import Agent, Runner, handoff, InputGuardrail, GuardrailFunctionOutput
215from pydantic import BaseModel
216 
217class SafetyCheck(BaseModel):
218    is_safe: bool
219    reason: str
220 
221async def content_safety(ctx, agent, input_text):
222    result = await Runner.run(
223        Agent(name="Safety", instructions="Check if input is safe. No PII."),
224        input_text,
225        context=ctx,
226    )
227    output = SafetyCheck.model_validate_json(result.final_output)
228    return GuardrailFunctionOutput(
229        output_info=output, tripwire_triggered=not output.is_safe
230    )
231 
232billing_agent = Agent(
233    name="Billing Agent",
234    instructions="You handle billing inquiries. Be precise with numbers.",
235    tools=[lookup_invoice, process_payment],
236)
237 
238refund_agent = Agent(
239    name="Refund Agent",
240    instructions="You process refund requests. Always verify the order first.",
241    tools=[lookup_order, issue_refund],
242)
243 
244triage_agent = Agent(
245    name="Triage Agent",
246    instructions="Route the customer to the right specialist. "
247                 "Ask clarifying questions if needed.",
248    handoffs=[billing_agent, refund_agent],
249    input_guardrails=[InputGuardrail(guardrail_function=content_safety)],
250)
251 
252result = await Runner.run(triage_agent, "I need a refund for order #4521")
253print(result.final_output)
254# The triage agent routes to refund_agent, which processes the refund
255```
256 
257### Silnye storony
258 
259- Chistyi pattern peredach  -  estestvennyi dlya workflow routinga/triazha
260- Guardrails rabotayut parallelno s vypolneniem (fail-fast, ne blokiruyushchie)
261- Vstroennaya panel tracing dlya otladki
262- Nesmotrya na nazvanie, podderzhivaet modeli ne ot OpenAI
263- Minimalnaya abstraktsiya  -  legko ponyat i rasshirit
264 
265### Slabye storony
266 
267- Menee zreloe upravlenie sostoyaniem, chem u LangGraph
268- Net vstroennoi postoyannosti ili checkpointinga
269- Ekosistema stroronnikh instrumentov menshe
270- Dizain, orientirovannyi na peredachi, mozhet ne podkhodit dlya kazhdoi arkhitektury
271 
272### Tseny
273 
274Open-source (MIT). Vy platite za token za lyubuyu ispolzuemuyu model.
275 
276## Claude Agent SDK: Razrabotchik
277 
278Claude Agent SDK prinimaet drugoi podkhod: vmesto opredeleniya workflow ili rolei vy dayote agentu **nabor instrumentov i pozvolyaete emu samomu razobr atsya, kak vypolnit zadachu**. On ispolzuet tot zhe avtonomnyi tsikl, kotoryi pitaet Claude Code  -  chitat, deistvovat, proveryat, iterirovat.
279 
280```mermaid
281graph TD
282    Prompt[User Prompt] --> Loop[Autonomous Agent Loop]
283    Loop --> Reason[Reason about next step]
284    Reason --> Act[Execute tool]
285    Act --> Verify[Check result]
286    Verify -->|not done| Loop
287    Verify -->|done| Output[Final output]
288```
289 
290### Klyuchevye kontseptsii
291 
292- **query()**: glavnaya tochka vkhoda, zapuskayushchaya tsikl agenta
293- **Vstroennye instrumenty**: Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch
294- **Polzovatelskie instrumenty cherez MCP**: opredelenie instrumentov kak in-process MCP serverov
295- **Sub-agents**: spetsializirovannye agenty, kotorym roditel mozhet delegirovat
296- **Sessions**: podderzhanie konteksta mezhdu neskolkimi vzaimodeistviyami
297 
298### Primer koda
299 
300```typescript
301import { tool, createSdkMcpServer, query } from "@anthropic-ai/claude-agent-sdk";
302import { z } from "zod";
303 
304const searchDocs = tool(
305  "search_docs",
306  "Search the internal documentation for relevant information",
307  { query: z.string().describe("Search query") },
308  async ({ query }) => {
309    const results = await vectorStore.similaritySearch(query, 5);
310    return {
311      content: [{ type: "text", text: results.map(r => r.pageContent).join("\n\n") }],
312    };
313  }
314);
315 
316const docsServer = createSdkMcpServer({
317  name: "docs",
318  version: "1.0.0",
319  tools: [searchDocs],
320});
321 
322for await (const message of query({
323  prompt: "Find how authentication works in our system and write a summary",
324  options: {
325    mcpServers: { docs: docsServer },
326    allowedTools: ["Read", "Glob", "Grep", "mcp__docs__search_docs"],
327  },
328})) {
329  if (message.type === "result" && message.subtype === "success") {
330    console.log(message.result);
331  }
332}
333```
334 
335### Silnye storony
336 
337- Pervoclassnaya integratsiya MCP  -  podklyuchaites k lyuboi ekosisteme MCP serverov
338- Vstroennye instrumenty dlya failovykh operatsii, terminala i dostupa v internet
339- Avtomaticheskoe szzhatie konteksta dlya bolshikh kodovykh baz
340- Parallelnye sub-agents dlya slozhnykh zadach
341- Tot zhe dvizhok, chto i Claude Code  -  proverennyi na realnykh workflow razrabotki
342 
343### Slabye storony
344 
345- Tolko modeli Claude  -  net podderzhki neskolkikh provaidderov
346- Bolee novyi freimvork s menshim soobshchestvom
347- Trebuet sredy Node.js dazhe dlya Python SDK
348- Menee yavnyi kontrol workflow po sravneniyu s LangGraph
349 
350### Tseny
351 
352Open-source. Standartnye tarify za tokeny Claude API. Managed Agents (khosting versiya): 0,08 $ za chas sessii plyus stoimost tokenov.
353 
354## Kogda vybrat kakoi
355 
356```mermaid
357graph TD
358    Start{What's your priority?}
359    Start -->|Full control over workflow| LG[LangGraph]
360    Start -->|Multi-agent collaboration| CA[CrewAI]
361    Start -->|Routing and triage| OA[OpenAI Agents SDK]
362    Start -->|Coding and file automation| CS[Claude Agent SDK]
363 
364    LG --> LGU[Complex stateful workflows\nConditional branching\nHuman-in-the-loop]
365    CA --> CAU[Team of specialized agents\nResearch + writing pipelines\nContent generation]
366    OA --> OAU[Customer service routing\nMulti-step handoffs\nInput validation]
367    CS --> CSU[Code generation and review\nFile-heavy automation\nMCP tool ecosystem]
368```
369 
370### Vyberite LangGraph, esli:
371- Vam nuzhenn tochnyy kontrol nad kazhdym shagom workflow
372- Vash sluchai ispolzovaniya vklyuchaet slozhnuyu uslovnuyu logiku i tsikly
373- Vy khotite vstroennuyu postoyannost i checkpointy human-in-the-loop
374- Vam nuzhno ispolzovat neskolko provaidderov LLM v odnom workflow
375 
376### Vyberite CrewAI, esli:
377- Vy khotite intuitivnuyu abstraktsiyu na osnove rolei
378- Vasha zadacha vklyuchaet neskolko agentov s raznymi spetsializatsiyami
379- Vam nuzhno, chtoby agenty sotrudnichali i peredavali drug drugu kontekst
380- Vy tsenite samoe bolshoe soobshchestvo i naibolshee kolichestvo vstroennykh integratsii
381 
382### Vyberite OpenAI Agents SDK, esli:
383- Vash osnovnoi pattern  -  marshrutizatsiya razgovorov k spetsialistam
384- Vam nuzhny guardrails, validiruyushchie vkhod/vykhod parallelno
385- Vy khotite maksimalno prostuyu abstraktsiyu s minimalnnym boilerplate
386- Vstroennyi tracing i nablyudaemost vazhny
387 
388### Vyberite Claude Agent SDK, esli:
389- Vashi agenty dolzhny chitat, pisat i vypolnyat kod
390- Vy khotite pervoclassnuyu integratsiyu s MCP serverami
391- Vam nuzhny avtonomnye agenty, kotorye iteriruyut i samoispravlyayutsya
392- Vy uzhe ispolzuete Claude i khotite naibolee glubokuyu integratsiyu
393 
394## Mozhno li kombinirovat freimvorki?
395 
396Da. Rasprostranyonnyi pattern  -  ispolzovanie odnogo freimvorka dlya orkestratsii i drugogo dlya otdelnykh agentov:
397 
398- **LangGraph** dlya obshchego grafa workflow
399- **CrewAI** dlya konkretnoigo uzla, trebuyushchego sotrudnichestva neskolkikh agentov
400- **Claude Agent SDK** dlya podzadach, svyazannykh s kodirovaniem, cherez MCP
401- **OpenAI Agents SDK** dlya triazha i routinga na storone klienta
402 
403Freimvorki ne isklyuchayut drug druga. Ispolzuite to, chto podkhodit dlya kazhdoi chasti vashei sistemy.
404 
405## Zaklyuchenie
406 
407Kazhdyi freimvork delaet chetkyy vybor:
408 
409- **LangGraph** optimiziruet kontrol  -  vy reshaete kazhdyi perekhod
410- **CrewAI** optimiziruet sotrudnichestvo  -  agenty rabotayut kak komanda
411- **OpenAI Agents SDK** optimiziruet prostotu  -  minimalnaya abstraktsiya, chistye peredachi
412- **Claude Agent SDK** optimiziruet avtonomiyu  -  daite emu instrumenty i pozvolte rabotat
413 
414Pravilnyi vybor zavisit ot vashego workflow, vashei komandy i vashego sushchestvuyushchego steka. Vyberite tot, kotoryi sootvetstvuet vashemu osnovnomu sluchayu ispolzovaniya, khorosho izuchite ego i podklyuchite drugie, kogda stoknyyotes s ikh silnoi storonoi.
415