When an AI agent learns, either the model, what it is shown, or its instructions change. Only the last two happen in Olano, and both are files you can read.

An AI agent that works across devices

Continuity comes from where the state lives, not from an app. One agent on a laptop, a phone, WhatsApp and Telegram, and why locking your phone stops nothing.

AI agent memory that persists: what survives what

Four stores do the job a features list calls memory. Which survive a model switch, a restart, a new phone, a new channel or deletion, and which you can edit.

AI agent platform lock-in: what you keep if the vendor disappears

Portability comes from formats, not promises. What an agent is on disk, what survives a vendor vanishing, what you would rebuild, and the questions to ask.

AI agent pricing: per seat, per usage or per deployment

A seat is a person and an agent is not one. The three ways AI agent platforms charge, what each shape optimises for, how to predict usage and compare quotes.

Portable agent skills: what travels between platforms and what does not

Skills written as folders of markdown and scripts travel between platforms; the tool names, credentials and helpers they lean on do not. Where the line falls.

Expose your agent as an MCP server or an A2A agent card

Most agent platforms are MCP clients. How an Olano agent becomes an MCP tool or an A2A agent card, why both are off by default, and what a caller can do.

Why your MCP connections keep breaking, and what a platform must do about it

Four ways an MCP OAuth connection dies - dropped refresh token, moved token endpoint, boot timeout, lapsed consent - and what a platform must do about each.

Self-hosted AI agents: the honest guide

The honest self-hosting guide: a machine that stays up, a stable address, a supervisor, secrets, backups, connector renewals, updates and the sandbox question.

How to teach an AI agent a new skill

A skill is a folder with a SKILL.md file, not a fine-tune. What goes in it, how one reaches an agent, the tools rule, and how to fork the bundled chart skill.

Where to start with AI in your business

For the majority of businesses that say they cannot see an applicable use case: how to find the first one without buying anything.

AI automation vs AI agents: which do you actually need?

Automation follows a path you drew. An agent decides the path. Knowing which problem you have saves you from buying the wrong shape.

How to improve efficiency with AI

Where the hours actually go in a small business, and which of them an agent can take without anyone losing control of the outcome.

AI for operations teams

The operating work - scheduling, chasing, reconciling, reporting - and what changes when an agent handles the first draft of each.

治理、安全与数据主权

如今适用于会采取行动的系统的规则、它们共同面临的攻击，以及这一切究竟运行在哪里。

The EU AI Act and AI agents: what changed in 2026

Transparency duties in force from 2 August 2026, enforcement powers live, high-risk obligations moved to 2027 and 2028 - and why agents are treated differently from chatbots.

US state AI laws in 2026: what to do while it settles

Twenty-nine states, a federal preemption push, and the four duties that recur in every version - so you can build once rather than fifty times.

Prompt injection and AI agents: a containment checklist

There is no reliable way to prevent it, so the defence is containment. What indirect injection looks like against a business agent, and eight controls that bound the damage.

Sovereign AI: where your agents actually run

Data, model calls, credentials and memory land in four different places. The six questions to put to a vendor in writing.

What self-hosting an AI agent really costs

The operational bill behind the infrastructure line: uptime, secrets, backups, connector renewals and the updates nobody scheduled.

研究与数据

2026 年关于采用率、失败率与投资回报的数据，每一项都注明发布机构，以及它实际统计的是什么。

AI agent statistics 2026: the figures worth quoting

Adoption, cancellation rates, productivity gains and security incidents, attributed and defined - because most of these numbers are quoted out of shape.

Why agentic AI projects get cancelled

More than 40% of them will be, on Gartner's forecast. The five failure modes behind that number, and what the survivors did differently.

How to measure AI ROI

Which numbers survive a budget review and which evaporate the moment somebody asks how they were calculated.

初创公司与市场开拓

写给刻意保持精简的团队：招人之前先交出哪些活、一次招聘的真实成本、一个智能体的真实成本，以及第一个企业客户会问什么。

AI agents for a startup: what to automate before you hire

The constraint at seed stage is founder attention, not headcount budget. Which jobs an agent carries, which it should not touch yet, and the order that works.

Hire or deploy? The 2026 US math

按 BLS 的数据，福利会在工资之外再加约 46%。比较的两边逐行写清楚。

What an AI agent actually costs to run

Four lines, and the subscription is the smallest. Why one agent quietly outspends six, and the cost line that is a person rather than compute.

Selling AI agents to enterprise buyers

The security questionnaire that arrives with your first serious customer, the four questions it asks about agents, and what to have ready before the call.

Is your AI agent actually working?

A broken agent answers promptly and sounds confident. Measuring task completion instead of response time, in three layers.

中国与东南亚

在中国平台与海外渠道之间工作：千问与混元作为模型，企业微信、钉钉、飞书作为渠道，以及一家从中国走向东南亚的公司真正需要什么。

Chinese companies expanding to Southeast Asia

Customers on WhatsApp and Telegram, the home team on WeCom - one deployment connecting both ends, with the compliance seen from both sides.

Running a Qwen-native agent team on Olano

Qwen as a first-class provider: one key, a model per role, the China and international endpoints, and open weights run locally.

Connecting AI agents to WeCom, DingTalk and Feishu

All three built in. How each connects, what it needs, and whether the agent should speak in a group.

Hunyuan and Tencent TokenHub on Olano

通过 OpenAI 兼容路线接入 Hunyuan（混元）：两种设置方式，以及各自什么时候值得用。

Wukong, Qianfan, Hunyuan agents and Olano: where each fits

A map by use case rather than a ranking, dated, and how the four coexist.

智能体如何协作

什么时候该把一个智能体拆成多个、它们之间如何交接工作，以及让它们触达自有部署之外的协议。

How many AI agents does a business need?

Fewer than you would expect. Add an agent when a job is genuinely distinct, not when it is large - more agents mean more supervision, not more capacity.

Multi-agent vs single agent: when to split

The case for several narrow agents over one that has accumulated every connector in the building - operationally and after a bad day.

智能体之间派活

Peers and subagents, what each is for, and why visibility and permission are deliberately separate settings.

What is agent orchestration?

The term, stripped of the marketing: who decides what runs next, and what breaks when nobody does.

MCP and A2A: what open agent protocols buy you

A2A reached v1.0 in April 2026 with 150+ organisations behind it. What each protocol does, and the part of lock-in neither of them solves.

自我改进的智能体

智能体在工作上变得更好意味着什么、不意味着什么，以及必须先到位的防护措施。

Recursive self-improvement, explained honestly

What the term means, what it does not mean, and which parts of it are actually shipping in products you can buy today.

Do AI agents learn from experience?

Sometimes, in a narrower sense than the phrase implies. What changes between Monday and Friday, and what stays exactly the same.

The four safeguards a self-improving agent needs

If an agent can change its own instructions, four things have to be true first. None of them is a model capability.

When not to let an AI agent self-improve

The cases where a fixed, reviewed, boring agent is the correct answer - and how to tell you are in one.

实操手册

针对具体流程的分步搭建方法，每一套都自带人工审批和审计记录。

平台内部

每一次部署背后的架构——系统如何自我改进、如何记忆、如何在多个 AI 模型之间分派工作，而且始终有你的审批环节。

The self-improvement engine: five cycles that turn daily work into sharper memory, skills and procedures on a cadence you control - Fully Autonomous by default, with a rewrite of an agent's own instructions always waiting for your approval.

AI 智能体是怎么记住事情的

记忆栈：不丢的对话记忆、提炼过的长期记忆、扎根于你自己文档的知识库，还有技能——全都在你专属的隔离环境里。

AI 智能体是怎么学会新技能的

反复做的活儿会被写成流程；对你装好的技能做改进走的是副本，从不覆盖原件；一旦哪份变差了，保全检查会把它隔离起来。

多模型 AI 智能体

每个任务都交给最合适的模型，可选 18+ 家 LLM 供应商——支出上限是硬的，自带密钥不额外收费，也不把你锁死在某一家。

对比

与通用 AI 工具、聊天机器人搭建工具、智能体平台和自建方案的并排对比，以及各自适合的场景。

客户故事

真实企业每天怎么用智能体，由他们自己讲。

一套连着库存的 WhatsApp 系统

一家摩托车配件商在自建库存 API 上跑两个 WhatsApp 智能体——给客户实时报价、查库存、发照片，员工用手机就能更新库存。

给一家营销代理做招标情报

每天两次监控政府招标，按契合度打分，再以邮件加 PDF 报告的形式送出：「这些报告太棒了，非常感谢！」

不如从一条流程开始

每个使用场景页面都写明具体搭法：智能体做什么、接哪些工具、审批卡在哪一步。

不确定该从哪条流程下手？

约一次咨询。我们梳理你最值钱的那条流程、给出报价，等你点头才动手——几天上线，不是几个月。

洽谈项目

让 AI 智能体真正干活的实用指南

精选

AI 智能体团队 vs 人类团队：一天的工作量对比

在人工审批之下自动处理客户咨询

开源

Singapore MCP：把新加坡官方数据装进你的 AI 助手

面向新加坡中小企业

新加坡中小企业资源见 olano.sg &rarr;

Olano Studio - 新加坡

指南

小企业用的 AI 智能体是什么？

要不要招一个「AI 实习生」？

如何挑选 AI 智能体平台

小企业最值得上的几条 AI 流程

Guide: always-on AI agents

Guide: personal vs business AI agents

Guide: what stops when you close your laptop

Agent memory is not fine-tuning: what "learning" actually means