uid=CeX1rsJL image ""

level="1" ✓uid=g836ZGlP StaticText "State of Agent Engineering" ✓uid=7qhpxd_b StaticText "We surveyed over 1,300 professionals — from engineers and product managers to business leaders and executives — to uncover the state of AI agents. Let’s dive into the data and break down how AI agents are being used (or not) today." ---- uid=qdUS4wwW link "" ---- uid=7YN92RK_ link "Introduction" ✓uid=DV2HQhmK StaticText "Introduction" ---- uid=prBKoyMc link "Insights" ✓uid=ogeiejoK StaticText "Insights" uid=JsqzLCsy link "Large enterprises are leading adoption" ✓uid=zgzcgZfK StaticText "Large enterprises are leading adoption" uid=eTYWVxen link "Leading agent use cases" ✓uid=PlVaBOx4 StaticText "Leading agent use cases" uid=xWN4IcNk link "Biggest barriers to production" ✓uid=gW3_uqAW StaticText "Biggest barriers to production" uid=0lTQ9nab link "Observability for agents" ✓uid=X5mGiutY StaticText "Observability for agents" uid=JwCvsGYf link "Evaluation and testing for agents" ✓uid=bhQNpWP6 StaticText "Evaluation and testing for agents" uid=WxWqK6It link "Model and tool landscape" ✓uid=76bX-Msx StaticText "Model and tool landscape" uid=nrAxvkwo link "What agents are being used daily?" ✓uid=tU_YOaJN StaticText "What agents are being used daily?" uid=duxcTvhg link "Methodology" ✓uid=kRLx4lHZ StaticText "Methodology" ---- uid=PDSLKcB3 heading "Introduction"
level="2" ✓uid=_j0HRUFj StaticText "Introduction" ✓uid=Qf4penSc StaticText "As we enter 2026, organizations are no longer asking whether to build agents, but rather how to deploy them reliably, efficiently, and at scale. We surveyed 1,300+ professionals to understand how they are evolving their AI agent use cases and handling agent engineering challenges." ---- ✓uid=7w0gu_le StaticText "Key findings:" list listitem ✓uid=4Mj0f77o StaticText "Production momentum is real," ✓uid=mKju5KCQ StaticText " with 57% of respondents having agents in production, with large enterprises leading in adoption" ---- listitem ✓uid=yHh6g5sZ StaticText "Quality is the production killer," ✓uid=2D6B8-jy StaticText " with 32% citing it as a top barrier. Meanwhile, cost concerns dropped from last year" ---- listitem ✓uid=Jj1AknQh StaticText "Observability is table stakes." ✓uid=5l8mUjNd StaticText " Nearly 89% of respondents have implemented observability for their agents, outpacing evals adoption at 52%." ---- listitem ✓uid=SkCXa77Q StaticText "Using multiple models is the norm." ✓uid=3Hx60nZ_ StaticText " OpenAI leads with their GPT models but Gemini, Claude, and open source models see significant adoption. Fine-tuning has not been widely adopted." ---- uid=BinFR2aF heading "Insights"
level="2" ✓uid=BILWqejL StaticText "Insights" blockquote ✓uid=4XpC2ES6 StaticText "What is agent engineering?" uid=vlUHbEuZ link "Agent engineering" ✓uid=K3hI73Ok StaticText "Agent engineering" ✓uid=MJ8fMus9 StaticText " is the iterative process of harnessing LLMs into reliable systems. Because agents are non-deterministic, we believe engineers need to rapidly iterate to refine and improve agent quality." ---- uid=Qe1tDIGj heading "Large enterprises are leading adoption"
level="3" ✓uid=ppVunNN1 StaticText "Large enterprises are leading adoption" paragraph ✓uid=AGMC1yIC StaticText "More than half" ✓uid=XjdEN4KH StaticText " of respondents surveyed (57.3%) now have agents running in production environments, with another 30.4% actively developing agents with concrete plans to deploy them." ---- ✓uid=OGod2fYZ StaticText "This marks clear growth from last year’s survey, where 51% reported having agents in production. Organizations are moving past the proof-of-concept stage into production — the question for most organizations is no longer “if” they will ship agents but “how” and “when”." ---- ✓uid=TtB-wfr6 StaticText "What changes at scale?" ✓uid=YZGr1Y6h StaticText "Of 10k+ size orgs, 67% had agents in production, with 24% actively developing with plans for production— versus for <100 size orgs, 50% had agents in production with 36% actively developing them. This suggest larger organizations are moving faster from pilots to durable systems, perhaps driven by greater investment in platform teams, security, and reliability infrastructure." ---- uid=RIR6Ydba heading "Leading agent use cases"
level="3" ✓uid=n8OMh-dF StaticText "Leading agent use cases" paragraph ✓uid=8G4z6ElF StaticText "Customer service emerged as the most common agent use case (26.5%)," ✓uid=PTc4RQfA StaticText " with research & data analysis close behind (24.4%). Together, these two categories represent more than half of all primary agent deployments." ---- ✓uid=4cCbA9iv StaticText "The strong showing of customer service suggests a shift toward teams putting agents directly in front of customers, not just using them internally. At the same time, agents continue to deliver clear value internally, with 18% citing using agents for internal workflow automation to boost employee efficiency." ---- ✓uid=aZORrybv StaticText "The popularity of research & data analysis use cases further reinforces where agents shine today: synthesizing large volumes of information, reasoning across sources, and accelerating knowledge-intensive tasks." ---- paragraph ✓uid=u17Z0fYk StaticText "Notably, we saw a " ✓uid=IsmExAoR StaticText "greater spread of use cases" ✓uid=TKRsz3M4 StaticText " selected by respondents this year (respondents could only select one primary use case), so agent adoption may be diversifying beyond a narrow set of early applications." ---- ✓uid=gBt-9hXW StaticText "What changes at scale?" paragraph ✓uid=n3kyiNIs StaticText "Among organizations with " ✓uid=5wEYwqKK StaticText "10k+ employees" ✓uid=wiNcTv-r StaticText "internal productivity is the top use case (26.8%)" ✓uid=U4x1UHU9 StaticText ", with customer service (24.7%) and research & data analysis (22.2%) close behind. Larger enterprises may tend to focus first on driving efficiency across internal teams before, or alongside, deploying agents directly to end users." ---- uid=4zwMv1To heading "Biggest barriers to production"
level="3" ✓uid=BQzo-ahd StaticText "Biggest barriers to production" paragraph ✓uid=65aaj4hm StaticText "Quality remains the biggest barrier to production" ✓uid=aTnwooo_ StaticText ", consistent with last year’s findings. This year, one third of respondents cited quality as their primary blocker. This encompasses accuracy, relevance, consistency, and an agent’s ability to maintain the right tone and adhere to brand or policy guidelines." ---- ✓uid=QHZ28yBh StaticText "Latency has emerged as second biggest challenge (20%). As agents move into customer-facing use cases like customer service and code generation, response time becomes a critical part of the user experience. This also reflects the tradeoff between quality and speed for teams, as more capable, multi-step agents can deliver higher quality outputs but often with slower responses." ---- paragraph ✓uid=HjX5yT61 StaticText "In contrast, " ✓uid=lJPYIhwl StaticText "cost is less frequently cited as a concern than in previous years" ✓uid=tHoTm-ya StaticText ". Falling model prices and improved efficiency appear to have shifted attention away from raw spend, with organizations prioritizing making agents work well and fast." ---- ✓uid=Fmbpn9u- StaticText "What changes at scale?" ✓uid=yDQbdQsy StaticText "Among enterprises (2k+ employees), quality remains the top blocker but security emerges as the 2nd largest concern, cited by 24.9% of respondents — surpassing latency, which is more commonly a challenge for smaller organizations." ---- paragraph ✓uid=abU-lxva StaticText "For orgs with 10k+ employees, write-in responses pointed to" ✓uid=NEz1GPZ0 StaticText " hallucinations and consistency of outputs" ---- ✓uid=COKOjRd8 StaticText " generated by agent as the biggest challenge in " uid=5kjsdUNu link "ensuring agent quality" ✓uid=z9SyLz5H StaticText "ensuring agent quality" ✓uid=DKgFYpLd StaticText ". Many also cited ongoing difficulties with " ---- uid=Jwaq9Rr9 link "context engineering" ✓uid=XuiCUqF2 StaticText "context engineering" ✓uid=_1RA-h5Z StaticText " and managing context at scale." ---- uid=xHVOfh0p heading "Observability for agents"
level="3" ✓uid=73eN0xNT StaticText "Observability for agents" paragraph ✓uid=_nJA1van StaticText "The ability to trace through multi-step reasoning chains and tool calls has become " ✓uid=rFSNsF95 StaticText "table stakes" ✓uid=bxln3TcU StaticText " for agents. " ---- ✓uid=4sQPNwGt StaticText "89% of organizations" ✓uid=rmZNgh10 StaticText " have implemented some form of observability for their agents, and 62% have detailed tracing that allows them to inspect individual agent steps and tool calls." ---- paragraph ✓uid=tiahhwlZ StaticText "Adoption is even higher among respondents who already have agents in production:" ✓uid=D9BS3nWe StaticText " 94% have some form of observability in place, and 71.5% have full tracing capabilities. This speaks to a fundamental truth of agent engineering: without visibility into how an agent reasons and acts, teams can’t reliably debug failures, optimize performance, or build trust with internal and external stakeholders." ---- uid=XGdkF6gb heading "Evaluation and testing for agents"
level="3" ✓uid=iA0founF StaticText "Evaluation and testing for agents" paragraph ✓uid=napTcOar StaticText "While observability is more widely adopted, agent evals are still catching up and gaining more awareness. Just over half of organization (52.4%) report running " ---- ✓uid=2bJFFnpb StaticText "offline evaluations" ✓uid=x8r3QTDW StaticText " on test sets, indicating that many teams see the importance of catching regressions and validating agent behavior before deployment. Adoption of " ---- ✓uid=flXLvN_S StaticText "online evals is lower " ✓uid=kpqXD-Lj StaticText "(37.3%) but growing as teams start to monitor real-world agent performance." ---- ✓uid=BGYbP-gl StaticText "Evaluation practices likely mature more for those who already have agents in production, as overall evals adoption is meaningfully higher (”not evaluating” drops from 29.5% to 22.8%). We also see a more organizations running online evals (44.8%), suggesting that once agents face real users, teams need to observe production data to detect issues in real time." ---- paragraph ✓uid=QcZcaPG7 StaticText "Most teams still start with " ✓uid=XoFx3oZK StaticText "offline evals" ✓uid=cFr_pxbL StaticText ", likely due to their lower barrier to entry and clearer setup, but many are layering approaches. Of the organizations running any evals, nearly a quarter combine " ---- ✓uid=7_wSNhUV StaticText "both offline and online evaluations." ---- paragraph ✓uid=mx7tpZQu StaticText "These organizations running evals also " ✓uid=mhndvPWT StaticText "rely on a mix of human and automated approaches for evaluation" ---- ✓uid=YtS--4bq StaticText ", implementing LLM-as-judge for breadth and human review for depth. More broadly, human review (59.8%) remains essential for nuanced or high stake situations, while LLM-as-judge approaches (53.3%) are increasingly used to scale assessments of quality, factual accuracy, and guideline adherence. In contrast, traditional ML metrics like ROUGE and BLEU see limited adoption. These may be less suitable for open-ended agent interactions where multiple valid responses exist." ---- uid=5WH6gnAw heading "Model and tool landscape"
level="3" ✓uid=J0ds6YHD StaticText "Model and tool landscape" paragraph ✓uid=PvOl52lw StaticText "OpenAI models dominate adoption, but " ✓uid=Gl0Wzutt StaticText "few teams are betting on a single provider." paragraph ✓uid=l6KMqvgW StaticText "More than two-thirds of organizations report using OpenAI’s GPT models, yet model diversity is the norm — with o" ---- ✓uid=kJtdNhia StaticText "ver three-quarters using multiple models in production or development." ✓uid=4F5qXPUe StaticText " Teams increasingly route tasks to different models based on factors like complexity, cost, and latency rather than going for platform lock-in." ---- ✓uid=GQhhNf_u StaticText "Despite the convenience of commercial APIs, running models in-house remains an importance strategy for many organizations. A a third of orgs report investing in the infrastructure and expertise required to deploy their own models. This open source model adoption may be driven by high-volume cost optimization, data residency and sovereignty requirements, or regulatory constraints in sensitive industries." ---- ✓uid=HIP0OmR3 StaticText "At the same time, fine-tuning remains specialized, not standardized. A majority of organizations (57%) are not fine-tuning models, instead relying on base models combined with prompt engineering and RAG. Since fine-tuning requires significant investment in data collection, labeling, training infrastructure, and ongoing maintenance, it seems to be reserved for mostly high-impact or specialized use cases." ---- uid=C79J4K1h heading "What agents are being used daily?"
level="3" ✓uid=T0bS9KGI StaticText "What agents are being used daily?" paragraph ✓uid=Vw0Doeja StaticText "When we asked, " ✓uid=PpBPfTHp StaticText "“What agents do you use the most in your day-to-day?”" ✓uid=nEZy3Rwc StaticText " a few clear patterns emerged from the write-in responses." ---- ✓uid=mxhiDLy4 StaticText "1. Coding agents dominate daily workflows." paragraph ✓uid=6vY0Y85p StaticText "By far the most commonly mentioned agents were coding assistants. Respondents repeatedly cited tools like " ---- ✓uid=sGWxFk0h StaticText "Claude Code, Cursor, GitHub Copilot, Amazon Q, Windsurf, and Antigravity" ✓uid=nZ6N25Tr StaticText " as part of their everyday development loop, whether it was for code generation, debugging, test creation, or navigating large codebases." ---- ✓uid=w8qCWCWH StaticText "2. Research & deep research agents are the next most used." paragraph ✓uid=acunHfQ- StaticText "The second most common pattern was " ✓uid=Vw9RypRY StaticText "research and deep research agents" ✓uid=KiZaLsGt StaticText " powered by ChatGPT, Claude, Gemini, Perplexity, and similar tools. These agents are used to explore new domains, summarize long documents, and synthesize information across sources. They’re often used as a companion to coding agents in the same workflow." ---- ✓uid=NJegbYZX StaticText "3. Custom agents built on LangChain and LangGraph are also popular." paragraph ✓uid=cSHxHesM StaticText "A third, distinct cluster of answers pointed to " ✓uid=hk2h5WEF StaticText "custom agents," ✓uid=udt-oaRw StaticText " with many respondents building on LangChain and LangGraph. Respondents described internal agents for things like QA testing, internal knowledge-base search, SQL/text-to-SQL, demand planning, customer support, and workflow automation." ---- ✓uid=aFsMn1ca StaticText "A meaningful minority also noted that they don’t yet use agents beyond LLM chat or coding assistance, underscoring that while agent usage is widespread, broader “agentic everything” is still in its early innings." ---- uid=adtEc611 heading "Methodology"
level="2" ✓uid=BAT0KI4E StaticText "Methodology" ✓uid=M9ymzi2q StaticText "The insights from this report were aggregated from a public survey that we ran for 2 weeks from Nov 18th-Dec 2nd, 2025. We received 1340 responses. Here’s some demographic data:" ---- ✓uid=IKT9Uy22 StaticText "Top 5 industries:" ✓uid=2q2wHJlq StaticText "Technology (63% of respondents)" ✓uid=kceYCjS- StaticText "Financial Services (10% of respondents)" ✓uid=17QkGihk StaticText "Healthcare (6% of respondents)" ✓uid=-4Rs8ibx StaticText "Education (4% of respondents)" ✓uid=h1qCMqYf StaticText "Consumer goods (3% of respondents)" ✓uid=0dO6yDpT StaticText "Manufacturing (3% of respondents)" ✓uid=EIt8DoNF StaticText "Company size:" ✓uid=lfgTtktU StaticText "<100 people (49% of respondents)" ✓uid=Id39z2VO StaticText "100-500 people (18% of respondents)" ✓uid=OAHLSVzH StaticText "500-2000 people (15% of respondents)" ✓uid=ALH0QDq2 StaticText "2000-10,000 people (9% of respondents)" ✓uid=VvfQgkHK StaticText "10,000+ people (9% of respondents)" ---- uid=ZxARzu6B heading "Check out relevant resources"
level="2" ✓uid=8kYVOuB0 StaticText "Check out relevant resources" ---- uid=2Q1-SKe1 link ""Agent Engineering: A New Discipline" Blog" uid=yTTKWxrs heading ""Agent Engineering: A New Discipline" Blog"
level="3" ✓uid=ArbjdF2- StaticText ""Agent Engineering: A New Discipline" Blog" uid=etg1OGyM link "LangSmith Evaluations Guidebook" uid=82lji5r_ heading "LangSmith Evaluations Guidebook"
level="3" ✓uid=CnDPnvXe StaticText "LangSmith Evaluations Guidebook" uid=sK_Lj57H link "LangSmith Observability" uid=lEOTFH2p heading "LangSmith Observability"
level="3" ✓uid=H4KwLU_B StaticText "LangSmith Observability" ---- ✓uid=jVvyHMOr StaticText "Products" ---- uid=2HfraiMO link "LangChain" ✓uid=kKBQ8RvV StaticText "LangChain" uid=-O-ee2ni link "LangGraph" ✓uid=MPeh9iLX StaticText "LangGraph" uid=vJWk0bbY link "LangSmith Observability" ✓uid=OScoJaHS StaticText "LangSmith Observability" uid=vmY0FHkb link "LangSmith Evaluation" ✓uid=iu8PGOAe StaticText "LangSmith Evaluation" uid=HhQWKO0Z link "LangSmith Deployment" ✓uid=t0Js1qJF StaticText "LangSmith Deployment" ---- ✓uid=PkboFtdA StaticText "Resources" ---- uid=6v8MpoJE link "Guides" ✓uid=AUn4bey5 StaticText "Guides" uid=826TmWyk link "Blog" ✓uid=C2Wlbpxw StaticText "Blog" uid=uvMMzq1_ link "Customer Stories" ✓uid=wUfEdbDS StaticText "Customer Stories" uid=k_FEVbpP link "LangChain Academy" ✓uid=7vDAQL6y StaticText "LangChain Academy" uid=HL0LsQu0 link "Community" ✓uid=ZghSkSMh StaticText "Community" uid=4ZbjPZJl link "Events" ✓uid=2RjpskRx StaticText "Events" uid=fNIKG2m1 link "Changelog" ✓uid=DUvBRZjo StaticText "Changelog" uid=y9me7syg link "Docs" ✓uid=NYzExz10 StaticText "Docs" uid=FjIplGg5 link "Support" ✓uid=t6blgDEy StaticText "Support" ---- ✓uid=nbl3TJqr StaticText "Company" ---- uid=aleAxAEW link "About" ✓uid=0FYEoIsu StaticText "About" uid=bzN7ZlVa link "Careers" ✓uid=j7ctX1_s StaticText "Careers" uid=NKsF4s8W link "X" uid=iuQbaLzM link "LinkedIn" ✓uid=IBbc_cMl StaticText "LinkedIn" uid=PkhyvjtQ link "YouTube" ✓uid=tbV8Qz5U StaticText "YouTube" uid=qGODt14Q link "Marketing Assets" ✓uid=s61eJBTv StaticText "Marketing Assets" uid=xXxol1nL link "Security" ✓uid=4J6U8o0B StaticText "Security" ---- ✓uid=9BcAjkQo StaticText "Sign up for our newsletter to stay up to date" ---- uid=zEi6_T_s textbox "your email..." ---- uid=bbXCLIWA link "All systems operational" ✓uid=b4LYOH0c StaticText "All systems " ✓uid=T-upAH9T StaticText "operational" ---- uid=jFjrb4d6 link "Privacy Policy" ✓uid=AjwzeEQT StaticText "Privacy Policy" uid=UubNFKjZ link "Terms of Service" ✓uid=9kWY7jSZ StaticText "Terms of Service"

总结此研究

Try it yourself