广告
加载中

对话火山引擎张鑫:Agent平台的主战场不在开发,在“用人”

张帅 2026-06-26 10:18
张帅 2026/06/26 10:18

邦小白快读

EN
全文速览

本文核心分享了当前企业Agent落地的行业现状与核心趋势,核心干货如下:

1. 当前中等规模且较早探索Agent的企业,内部平均已经有大约三百个Agent在运行,但绝大多数企业说不清Agent的数量、使用人群、成本消耗和实际效果,还没有建立以Agent为核心的管理体系。

2. 现在零代码等技术已经让搭建Agent不再是门槛,企业落地Agent的核心卡点已经从“开发”转移到“使用和管理”,普遍存在三个痛点:家底不清、成本黑洞、缺乏绩效度量,同时还存在端到端流程集成难的问题。

3. Agent平台整体市场规模预计在500亿到千亿之间,真正的爆发期大概率出现在2026年底到2027年,未来付费模式会逐步转向按业务效果收费,行业增长空间很大。

品牌商布局AI Agent落地、推进业务智能化可以获得以下核心干货:

1. 当前Agent已经从尝鲜性质的创新产品,转变为深度融入企业核心生产和业务场景的数字员工,品牌布局AI不能只停留在单点开发Agent,要转向搭建全链路的Agent管理体系,才能真正实现全业务提效增值。

2. 品牌落地Agent大概率会遇到三大共性痛点:说不清内部有多少Agent运行、无法统计Token成本消耗、无法评估Agent绩效,对此可以参考成熟方案,搭建统一的Agent管理入口,搭配通用开箱即用应用和专属业务Agent,形成联动的产品矩阵。

3. 价值衡量方面,领先企业已经开始将Agent投入从软件预算转向人力预算,品牌可以参考这种模式,从人力成本节省、业务效果增值两个维度评估Agent价值,比单纯看Token消耗更能体现实际价值。

做AI相关业务的卖家可以从中得到以下机会提示与风险参考,核心内容如下:

1. 行业增长趋势明确,据IDC数据,Agent开发平台市场2024年约四五亿,2025年增长到约20亿,年增速达三四倍,整体市场规模未来将达到500亿到千亿,真正爆发期在2026年底到2027年,赛道增长空间十分广阔。

2. 当前行业核心痛点已经转移,企业落地Agent的卡点不再是开发,而是使用和管理,多数企业存在家底不清、成本失控、绩效无法度量三大痛点,还有端到端流程集成难的问题,卖家可以针对这些痛点布局产品和服务,抢占市场机会。

3. 需要注意的风险是:当前技术范式迭代快,企业容易陷入技术选型焦虑,同时企业技术认知跟不上技术门槛下降的速度,落地Agent需要配套的落地咨询服务,卖家可以配套布局相关服务提升竞争力。

工厂推进数字化转型、布局AI Agent可以获得以下核心启示:

1. 当前Agent技术已经足够成熟,零代码工具让搭建Agent的门槛大幅降低,工厂也可以低成本尝试搭建适配自身生产、采购、管理等不同需求的专属Agent,开发层面已经不存在难以突破的门槛。

2. 工厂布局Agent不能只停留在单点提效,如果只优化单个环节,上下游流程还是人工卡点,单个环节的效率提升对整体生产没有实质意义,必须重视端到端全流程的集成打通,才能真正发挥Agent的价值。

3. 工厂引入多Agent之后,要建立统一的管理体系,像管理员工一样管理Agent,清楚掌握Agent运行数量、成本消耗、绩效情况,及时淘汰低效Agent,放大高效Agent的价值;还可以结合工厂业务特性,强管控稳态任务用Workflow,发散性探索场景用Agentic Loop,两种范式灵活搭配提升效率。

面向企业提供AI服务的服务商可以得到以下行业干货,核心内容如下:

1. 当前行业发展趋势已经发生根本性改变,Agent开发技术门槛已经大幅降低,造Agent不再是难事,客户的核心焦虑已经从“怎么开发Agent”转变成“怎么用好管好Agent”,行业竞争主战场已经从开发转移到落地运营管理。

2. 当前客户的核心痛点可以总结为三点:一是家底不清,说不清内部有多少Agent在运行、哪些人在使用;二是成本黑洞,无法精准统计每天的Token消耗成本;三是缺乏度量标准,无法评估Agent的实际绩效,不知道该下架低效Agent还是加大投入,此外还有端到端流程集成难的问题。

3. 服务商可以布局的方向包括:打造覆盖开发、运行、消费、管理全四个域的全生命周期管理能力;布局Harness层满足企业可控、可靠、可审计等六大核心需求;将传统交付人员转型为前沿部署工程师FDE,深入业务提取隐性知识,可持续服务客户同时反哺自身产品迭代。

做Agent平台的厂商可以从文中得到以下产品、运营、商业化的参考经验,核心内容如下:

1. 当前企业对Agent平台的核心需求已经发生变化,从只要求提供开发能力,转向要求提供全生命周期的统一管理能力,平台只布局开发域已经无法形成差异化竞争力,需要延伸到运行、消费、管理全链路。

2. 产品策略上,可以参考火山引擎的调整,将原来“1+N+X”体系中的“1”从统一交互入口调整为统一管理入口,通过统一入口管理所有Agent,N定位为开箱即用的通用应用,X支持企业搭建专属业务Agent,形成联动的产品矩阵满足不同需求。

3. 核心竞争壁垒集中在三层:一是做好Harness层,满足六大核心需求,保障模型能力在生产环境稳定发挥;二是适配企业敏态稳态需求,让Workflow和Agentic Loop共存互补;三是复用企业已有软件生态,改造调用方式而非重新开发,降低企业落地成本。商业化方面未来会逐步转向按业务效果收费,平台需要提前适配。

研究AI Agent产业的研究者可以得到以下关于产业新动向、新问题、新商业模式的干货内容:

1. 当前产业已经出现根本性的新动向,Agent已经从企业尝鲜的创新产品,转变为深度融入核心生产场景的数字员工,整个市场的竞争主战场已经从开发域转移到应用和管理域,全生命周期管理成为平台差异化竞争的核心方向。

2. 当前产业发展存在的新问题包括:技术范式迭代速度快,企业普遍存在技术选型焦虑;技术开发门槛不断下降,但企业的技术认知没有同步提升,多数企业还不具备搭建复杂业务Agent的能力;现有的按Token收费模式无法体现不同场景的价值差异,企业的组织和流程也还没有做好Agent规模化落地的准备。

3. 产业未来发展方向明确:整体市场规模将达到500亿到千亿,真正爆发期在2026年底到2027年;商业模式会逐步从按资源收费、按Token收费,升级为按业务效果收费;竞争壁垒集中在Harness层、运行范式理解、软件生态重构三个层面。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article shares key insights on the current industry status and core trends of enterprise AI Agent deployment:

1. Medium-sized companies that started exploring AI Agent early now run an average of around 300 Agents internally. However, the vast majority of enterprises cannot clearly account for their Agent inventory, user base, cost consumption or actual business impact, and have not yet established an Agent-centric management system.

2. Thanks to technologies like no-code development, building Agents is no longer a major barrier. The core bottleneck for enterprise adoption has shifted from "development" to "usage and management". Three common pain points are unclear inventory, uncontrolled costs and lack of performance measurement, alongside persistent difficulties in end-to-end process integration.

3. The overall market size of AI Agent platforms is projected to reach between 50 billion and 100 billion RMB. The industry will most likely enter its true growth explosion phase between late 2026 and 2027. Going forward, the pricing model will gradually shift to charging based on business outcomes, leaving significant room for long-term industry growth.

This article shares key takeaways for brands looking to deploy AI Agents and advance business digitalization:

1. AI Agents have evolved from experimental innovation projects to digital workers deeply integrated into enterprises' core production and business scenarios. Brands should not stop at just developing isolated Agents; instead, they need to build an end-to-end Agent management system to unlock efficiency gains and value creation across the entire business.

2. Brands commonly face three major pain points when deploying Agents: unclear inventory of running Agents, inability to track total Token cost consumption, and lack of framework to evaluate Agent performance. A proven solution is to build a unified Agent management portal, combine out-of-the-box general applications with custom business-specific Agents, and form an interconnected product matrix.

3. For value measurement, leading companies have already shifted Agent investment allocation from software budgets to human resources budgets. Brands can adopt this framework, and evaluate Agent value from two dimensions: human cost savings and business outcome improvement, which better reflects actual value than measuring Token consumption alone.

This article outlines key opportunity signals and risk references for sellers working on AI-related business:

1. The industry has a clear upward growth trajectory. According to IDC data, the Agent development platform market will reach approximately 400-500 million RMB in 2024, and grow to around 2 billion RMB in 2025, representing an annual growth rate of 300-400%. The overall long-term market size will reach 50 billion to 100 billion RMB, with the mass adoption boom arriving between late 2026 and 2027, creating enormous growth space for the sector.

2. The core industry bottleneck has shifted: the main challenge for enterprise adoption is no longer development, but usage and management. Most enterprises face three core pain points: unclear Agent inventory, uncontrolled costs, and inability to measure performance, in addition to end-to-end integration difficulties. Sellers can build products and services tailored to these pain points to capture early market share.

3. Key risks to note: The current technical paradigm is evolving rapidly, leaving enterprises vulnerable to technology selection anxiety. Meanwhile, enterprise technical knowledge has not kept pace with falling development barriers. Successful Agent deployment requires supporting implementation and consulting services, so sellers can add these complementary services to improve their competitive edge.

This article shares key insights for factories advancing digital transformation and deploying AI Agents:

1. Agent technology is now sufficiently mature, and no-code tools have drastically lowered the barrier to building Agents. Factories can now build custom Agents tailored to their specific needs in production, procurement, internal management and other areas at low cost, with no insurmountable barriers on the development side.

2. Factories should not stop at optimizing single isolated processes. Efficiency gains in a single环节 will not deliver material improvements to overall production if upstream and downstream processes still rely on manual work. Factories must prioritize end-to-end full-process integration to unlock the full value of AI Agents.

3. After deploying multiple Agents, factories need to establish a unified management system, managing Agents just like they manage human employees: keep clear track of Agent quantity, cost consumption and performance, eliminate low-efficiency Agents in time, and scale the value of high-performing Agents. Factories can also flexibly combine two paradigms based on their business characteristics: use Workflow for tightly controlled stable tasks, and Agentic Loop for open-ended exploratory scenarios to maximize overall efficiency.

This article shares key industry insights for AI service providers serving enterprise clients:

1. The industry has undergone a fundamental shift in development trends. Agent development barriers have fallen sharply, and building Agents is no longer difficult. Clients' core anxiety has shifted from "how to build an Agent" to "how to use and manage Agents well", and the main competitive battlefield has shifted from development to post-deployment operation and management.

2. Clients' core pain points can be summarized into three categories: first, unclear inventory, with no clear picture of how many Agents are running internally and who is using them; second, cost black holes, with no way to accurately track daily Token consumption costs; third, lack of measurement standards, with no way to evaluate actual Agent performance to decide whether to retire low-performing Agents or increase investment. End-to-end process integration is also a persistent pain point.

3. Promising areas for service providers to expand include: building full-lifecycle management capabilities covering development, operation, consumption and management domains; building out the Harness layer to meet enterprises' six core requirements of controllability, reliability, auditability and more; retraining traditional delivery staff into Frontier Deployment Engineers (FDE) who can extract tacit knowledge from deep within client business, enabling sustained client service and feeding back insights to improve product iteration.

This article shares product, operation and commercialization insights for Agent platform vendors:

1. Enterprise demand for Agent platforms has fundamentally changed: instead of only requiring development capabilities, enterprises now demand unified full-lifecycle management capabilities. Platforms that only focus on the development domain can no longer build differentiated competitive advantages, and need to extend their coverage across the full development, operation, consumption and management chain.

2. For product strategy, vendors can reference Volcano Engine's adjustment: it repositioned the "1" in its original "1+N+X" framework from a unified interaction portal to a unified management portal, which manages all Agents through one single entry. "N" refers to out-of-the-box general applications, while "X" enables enterprises to build custom business Agents, forming an interconnected product matrix that meets diverse needs.

3. Core competitive barriers fall into three layers: first, building a robust Harness layer that meets six core requirements to ensure stable performance of model capabilities in production environments; second, supporting both enterprise stable and agile needs by enabling coexistence and complementarity between Workflow and Agentic Loop; third, integrating with enterprises' existing software ecosystems by adapting calling methods rather than rebuilding systems from scratch, lowering enterprises' deployment costs. For commercialization, the industry will gradually shift to outcome-based pricing, so platforms need to prepare for this transition in advance.

This article shares insights on new industry trends, emerging problems and new business models for AI Agent industry researchers:

1. The industry has already seen fundamental new shifts: AI Agents have evolved from experimental innovation projects for enterprises to digital workers deeply integrated into core production scenarios. The entire industry's competitive focus has shifted from the development domain to the application and management domain, and full-lifecycle management has become the core direction for platforms to build differentiated competitive advantage.

2. Key new problems facing industry development include: rapid iteration of technical paradigms leading to widespread technology selection anxiety among enterprises; development barriers have continued to fall, but enterprise technical awareness has not kept pace, and most enterprises still lack the capability to build complex business Agents; the existing Token-based pricing model cannot reflect value differences across different scenarios, and enterprise organizational structures and processes are not yet prepared for large-scale Agent deployment.

3. The future direction of industry development is clear: the overall market size will reach 50 billion to 100 billion RMB, with the mass growth boom arriving between late 2026 and 2027; business models will gradually upgrade from resource-based and Token-based pricing to outcome-based pricing; and core competitive barriers will concentrate on three areas: the Harness layer, operational paradigm design, and software ecosystem reconstruction.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

火山引擎将如何回答企业Agent规模化落地的系统命题,时隔一年之后,钛媒体App和张鑫做了一次对话。

2026年中,一家中等以上规模且较早探索Agent的企业,内部有多少Agent在跑?平均值约为三百个。

厘清Agent的数量都是一个难以统计的答案,若继续追问,这些Agent具体是哪些人在用、烧了多少钱,效果具体如何,几乎没有企业能说清楚。这不是个例,而是现阶段的企业,并没有建立一个以Agent为核心的系统和组织。

换而言之,以前企业更多是单点的开发、使用Agent,随着Agent数量和质量的提升,企业越是深入使用Agent,就会越早发现另一重难题——Agent在企业的规模化落地,比预想中的复杂且繁琐。

火山引擎副总裁张鑫表示,Agent的效果绝对不是仅仅由开发态决定的,而是需要一个全域的、不断滚动的过程。“这些Agent烧了多少的Token,每天的绩效怎么样,哪些绩效不好该‘开除’,哪些绩效很好要‘升职’,企业需要一个统一管理的平台。”

这也是张鑫观察到的整个市场的根本性改变,多数人还在盯着怎么造Agent时,真正的战场已经转移了,Agent从一种尝鲜性质的创新产品,演变成深度融入企业所有核心生产和中长尾场景的数字员工,Agent平台的核心竞争力也随之变化。

今年以来,扣子3.0、TRAE企业版、HiAgent 3.0、AgentSphere四款产品的迭代和商业化落地明显提速,本质上就是为了应对这种趋势,火山引擎将如何回答企业Agent规模化落地的系统命题,时隔一年之后,钛媒体App和张鑫做了一次对话。

当造Agent不再是问题

过去一年,Agent开发平台的技术迭代速度超出所有人预期。工作流、Skills、Agentic Loop、Vibe Coding,新的构建方式层出不穷,零代码工具让搭建智能体变得越来越简单。

“企业里构建Agent已经不是门槛了,大家谁都能构建。”张鑫说,但“能造”和“造好”之间,隔着一道巨大的鸿沟。

张鑫走访大量企业后发现,企业的焦虑非但没有减少,反而加剧了。去年企业以为找到了路径,通过低代码、工作流去构建Agent,2026年春节开始,新的技术不断在推翻前面的范式,选择多了,企业反而更困惑,以前买的产品是不是过时,这是企业问得最多的问题。

焦虑的根源在于,企业发现真正的卡点不在“造”,而在“用”。大部分企业已经能做到局部提效,代码生成环节接一个TRAE或Claude Code,从一小时降到5分钟。但企业的端对端流程里,可能有十余个步骤,上下游的测试、发布、运维还是人工处理。

“如果你只解决其中一个AI步骤,上游下游还是卡着,中间缩短的那1小时没有任何意义,”张鑫说。这些端对端的系统集成,看起来是脏活累活,跟底层模型迭代相比没那么令人兴奋,但这确实就是真正做到大企业深水区,Agent落地绕不过的门槛。

更棘手的是,Agent造出来之后,没有一个清晰的管理工具。张鑫总结了三个具象的痛点:家底不清,企业到底有多少个Agent在跑,都是谁在用,回答不清楚;成本黑洞,每天烧了多少Token、产生了多少成本,很多企业内部没有统计;缺乏度量,怎么评估Agent的绩效,哪些该下架,哪些该加大投入。

这个判断直接影响了火山引擎的产品策略。去年提出的“1+N+X”体系里,“1”是想做统一交互入口,今年张鑫发现,每个产品都能部分承担AI工作台的入口角色,企业管理者依旧需要一个统一的管理入口,所以火山引擎对于“1”的定位,从统一交互入口变成了统一管理入口。

围绕“1”从统一交互入口变成统一管理入口的判断,可以看出,企业需要的从来不是几个孤立的Agent,而是一套面向不同角色和部门、彼此联动的产品矩阵。通过“1”的数字员工统一管理入口,将所有数字员工纳入统一体系经营与度量;“N”是多个开箱即用的智能化应用,承接企业通用需求;“X”是无限多个持续进化的业务智能应用,企业可基于TRAE、扣子、HiAgent持续创建、运行、观测、优化专属数字员工。

Agent开发之后的主战场

既然开发不再是壁垒,壁垒到底在哪?

张鑫把Agent的全生命周期拆成四个域:开发域搭建Agent;运行域通过Harness让Agent跑得更好,包括记忆管理、知识融合、上下文管理、Multi-Agent编排、意图识别;消费域通过好的交互和人机协同方式让员工使用Agent;管理域像管理数字员工一样管理Agent,做绩效考评、度量价值。

“Agent的效果绝对不仅仅由开发域决定,”他说,“在运行域,你的Harness做得够不够好,在消费域,人机协同的机制设计得好不好,Agent需设计好人在关键节点介入的机制,让最终效果达到100分。在管理域,能不能像管人一样管数字员工。”

由此可以看出,火山引擎已经不只在开发域和其他平台竞争,而是通过四个域的全生命周期管理形成差异化,这个判断的背后,是三个层面的竞争壁垒。

第一层壁垒是Harness。模型能力越来越强,会吃掉多少中间层,是一个被反复讨论的话题。

张鑫的判断是,Harness不会被吃掉。“Harness解决六个维度的问题:可控、可靠、可审计、可规模化、全生命周期管理、价值可度量。以当前Next Token Prediction的技术架构,总需要外部的交叉验证和行业知识。”

有些部分会被模型吸收,比如一步步控制Agent怎么走的机制,随着模型能力增强,会从Control转向“Context Not Control”,但那六个维度不会被替代。模型决定了智力的天花板,Harness决定了这个智力能不能在生产环境里安全、可控、可度量地发挥出来。

第二层壁垒是对运行范式的理解。很多人判断Workflow将被淘汰,Agentic Loop会成为主流。张鑫认为,两者需要共存,因其本质上都是对工作流流程的表达,只是载体不同。

企业内部里有“敏态”和“稳态”之分,强管控的任务需要Workflow保证可靠性,更发散的场景适合通过Agentic Loop探索,两者可以相互转化,一开始通过Agentic Loop探索,跑出最佳路径后沉淀为Workflow,效率更高、Token消耗更少。

第三层壁垒是对软件生态的重构,“软件日抛论有点脱离实际,”张鑫说,“企业Token成本在不断飙升,已有工具不去利用还要重新搞一套,非常不现实。”

譬如,CLI化不会吃掉软件,但会改变软件的交互入口,从给人用变成给Agent用,而软件背后的逻辑和数据不变,只是调用方式变了。

这三层壁垒加在一起,构成了开发域之外的竞争空间。与此同时,火山引擎也在探索前沿部署工程师(Forward Deployed Engineer,简称FDE)的实践,企业需要有人把Agent落地到具体业务之中。

张鑫表示,“FDE和传统交付有本质区别,传统交付追求一次性项目验收,FDE更追求可持续的Token消耗,打造AI Native产品,同时反哺产品。”

FDE最难的是把业务中的隐性知识提取出来,工程能力的门槛反而因AI Coding而降低,为此,火山引擎将离业务近的、有咨询能力的人员,转型为FDE,离实际业务场景效果比较近,FDE的价值才能更显性。

千亿市场,亟待爆发

对于AI软件市场,行业存在两种观点,一种是,大模型是如同互联网一般的基础设施,其上会长出繁荣的应用,另一种观点是,大模型本身会越来越多地“吃掉”软件本身。

按照IDC给出的数据来看,2024年Agent开发平台市场大约四五亿,2025年大约20亿,每年三四倍增速。火山引擎凭借HiAgent和扣子,分别以17.8%和19.3%的份额,位居2025年中国智能体开发平台私有化、公有云市场第一。

张鑫认为,Agent平台整体市场测算在500亿到千亿规模,真正爆发期可能在2026年底到2027年。

客观上,模型能力依旧有待突破,而且企业的组织没准备好、流程没打通、门槛尚不够低。

技术供应一侧不断降低门槛,企业的技术认知却没有同步跟上,零代码工具不等于实现零认知。以HiAgent为例,客户数是大几百家量级,和中国企业总量相比差得非常远。真正能搭建复杂业务场景工作流的企业还是少数。

张鑫表示,突破方向有三条,足够好的Baseline,让用户不必从零搭建;提供开箱即用的模板和Skills;通过Coding能力解决更复杂的任务,从大模型使用已有工具,变成大模型能按需创造新工具。

而一旦市场真正爆发,商业模式也会有所变化,第一层按资源收费(GPU卡时),第二层按Token收费,第三层按业务效果收费。目前行业主要还在第二层,但按Token收费存在一个“盲区”,同样100万Token,用来做纯聊天和用来做芯片设计,业务价值天差地别,目前按Token收费体现不出这种差别。

“到了第三层,看实际产出了多少可用代码、多少可用视频时长,就不需要纯看底层烧了多少Token,因为很多可能是无用Token。”张鑫说。

此外,一部分领先企业,开始将Agent从软件预算转向人力预算。衡量数字员工价值,最直观的是人力成本节省,更高维度是业务效果增值,审核从人工几小时变成秒级上线,客服质量标准化统一,这些维度共同构成了更高的Agent市场想象空间。

对于火山引擎的智能体平台产品和商业化目标,张鑫直言,“大原则不变,每一个产品都要做到市场占有率第一。但最终,大家还是会以拉动了多少Token来衡量产品是不是真正产生了价值。”

注:文/张帅,文章来源:钛媒体(公众号ID:taimeiti),本文为作者独立观点,不代表亿邦动力立场。

文章来源:钛媒体

广告
微信
朋友圈

这么好看,分享一下?

朋友圈 分享

APP内打开

+1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0