Resources
Practical guides to putting AI agents to work
Practical guides to deploying AI agents across WhatsApp, email, CRM, research, and marketing workflows. Every guide reflects how we build: start with one workflow, keep a human in the loop for anything outbound, and keep an immutable audit trail of every action.
Featured
Start here: what a supervised agent team actually gets done in a day, and the playbook for automation your customers feel - without a single unreviewed message leaving your business.
AI agent team vs human team: one day of work compared
One real day of mixed operating work - research, CRM prep, outreach, monitoring, support - priced honestly as human labour.
- 147 companies and 43 contacts researched, validated, and added to the CRM
- 80-140 human-equivalent hours, estimated from per-unit assumptions - S$5,000-S$25,000+ across execution models
- Where the speed advantage ends: the human approval boundary
Automate customer enquiries with human approval
The step-by-step playbook for letting an agent handle enquiries while your team approves every outbound reply.
- Trust levels 0-4, from Observer to Developer
- Per-category approval gates - outbound messages wait for a human by default
- An immutable audit trail on every action the agent takes
Built on LangChain
Olano runs on LangChain, LangGraph and Deep Agents. How the stack fits together, and what running it in production taught us about debugging it.
Powered by LangChain: how Olano is built
Which layer of the LangChain stack does what inside an Olano agent, and what Olano adds on top for a business.
Debugging LangChain agents: a field guide
Debug mode, LangSmith traces, stream modes, saved state, loop limits and fake models for reproducing a bug offline.
Debugging LangGraph checkpointers
Missing thread ids, memory lost on restart, event-loop errors, Postgres setup, encryption and deleting a thread properly.
Debugging LangChain streaming
Leaked tokens from models called inside tools, garbled replies, fused sentences and streams that must never be retried.
Switching LLM providers mid-conversation
Content blocks one provider writes and another rejects, unknown context windows and parameters that change name.
Open source
What we build and give away - free, MIT-licensed, and usable without an Olano account.
For Singapore SMEs
Singapore-specific guides - WhatsApp customer service, PDPA, grants, hiring math, and sector playbooks - now live on the Olano Studio site.
Singapore SME resources at olano.sg →
The full local library: AI agents for Singapore SMEs, WhatsApp AI customer service, AI grants (PSG, EDG, EDGE), PDPA and AI agents, the hire-vs-deploy math, and retail, F&B, and property playbooks.
Olano Studio - Singapore
The implementation arm: workflow discovery, system design, integration, dedicated deployment, and ongoing management for Singapore businesses.
Guides
Foundations: what AI agents are, where they fit in a small business, and how to choose a platform.
What is an AI agent for a small business?
A plain-English explainer: how an agent differs from a chatbot, and why your own team of agents is the more useful mental model.
Should you hire an AI intern?
The intern test: what the metaphor gets exactly right about supervised AI agents - and where it breaks, mostly in your favour.
How to choose an AI agent platform
A buying checklist built on what actually matters: channels, approvals, audit trails, spend caps, BYOK, and workspace isolation.
Best AI workflows for a small business
Five workflows worth automating first - customer service, inbox triage, research, CRM follow-ups, and social drafting - and how to pick one.
Guide: always-on AI agents
What 24/7 actually requires - five mechanisms, the question nobody asks in a demo, and four limits any honest vendor will name.
Guide: personal vs business AI agents
A consumer agent is issued to a person, and a business is not a person - the five things that follow from that.
Guide: what stops when you close your laptop
The overnight failure, diagnosed - seven things that need a live process, and the connector-expiry trap nobody expects.
Agent memory is not fine-tuning: what "learning" actually means
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.
Governance, security and sovereignty
The rules that now apply to systems that take actions, the attack they are all exposed to, and the question of where any of it runs.
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.
Research and data
The 2026 figures on adoption, failure and return - each with the organisation that published it, and a note on what it actually counts.
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.
Startups and go-to-market
For a team that is small on purpose: what to hand over before you hire, what a hire actually costs, what an agent actually costs, and the questions your first enterprise customer will ask.
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
Benefits add roughly 46% on top of wages on BLS figures. Both sides of the comparison written out, line by line.
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.
China and Southeast Asia
Working across the Chinese platforms and the overseas channels: Qwen and Hunyuan as models, WeCom, DingTalk and Feishu as channels, and what a company expanding from China into Southeast Asia actually needs.
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
Hunyuan through an OpenAI-compatible route: two ways to set it up, and when each is worth taking.
Wukong, Qianfan, Hunyuan agents and Olano: where each fits
A map by use case rather than a ranking, dated, and how the four coexist.
How agents work together
When one agent should become several, how they hand work to each other, and the protocols that let them reach outside your own deployment.
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.
Agent-to-agent delegation
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.
Self-improving agents
What it means for an agent to get better at its job, what that does not mean, and the safeguards that have to exist first.
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.
Playbooks
Step-by-step setups for specific workflows, each with human approval and an audit trail built in.
Automate customer enquiries with human approval
Handle enquiries automatically while a human approves every reply before it is sent - the approval-first setup end to end.
AI inbox triage and follow-ups
Put an agent on your shared inbox to triage what arrives, draft replies, and queue follow-ups for your approval.
AI agent competitor monitoring
Use scheduled tasks and proactive monitoring to watch competitors and deliver research digests to your team.
AI social media agent with an approval workflow
Draft posts with an agent, route every one through approval, and publish through connected tools like Buffer.
Inside the platform
The architecture behind every deployment - how the systems improve, remember, and route work across AI models, always with your approval in the loop.
Inside Olano Cortex
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.
How AI agents remember
The memory stack: persistent conversation memory, distilled long-term memory, a knowledge base grounded in your documents, and skills - in your own isolated environment.
How AI agents learn new skills
Recurring work becomes a written procedure; improvements to skills you installed are forked, never overwritten; and a preservation gate quarantines anything that degraded.
Multi-model AI agents
Each task routed to the model best suited to it across 18+ LLM providers - with hard spend caps, BYOK at no extra cost, and no vendor lock-in.
Comparisons
Side-by-sides with general-purpose AI tools, chatbot builders, agent platforms and DIY stacks - and where each one fits.
We asked Grok Bot to compare itself with Olano
xAI's AI teammate analysed cloud.olano.ai, compared it with itself and rated both. The prompt, the screenshot, the full reply - and why its most useful answer was about who each one serves.
Diary: AI agent team vs human team
One real day of operating work compared with its human-team equivalent - hours, Singapore-dollar costs, and where the human approval boundary sits.
Guide: AI agent vs chatbot
The plain-English explainer - what separates a scripted chatbot from an agent that takes approved actions across your systems.
Table: AI agent platform vs chatbot
The feature-by-feature side-by-side: chatbot builders versus an agent platform with approvals, audit history, and spending controls.
Guide: Olano vs ChatGPT for operations
The explainer - where ChatGPT fits, and where persistent agents on customer channels earn their keep.
Table: Olano vs ChatGPT for business
The detailed side-by-side of a general-purpose AI workspace and a hosted platform for agents on your business channels.
Compare: hosted AI agents vs DIY automation
Hosted, managed agents versus building your own stack - cost, maintenance, and where each approach fits.
Table: Olano vs Glimmer (Glints)
Two Singapore answers to the same question - an AI teammate in your staff chat, against agents on your customer channels.
Table: Olano vs Lindy
Self-serve credit-metered AI employees versus a managed, isolated deployment - and where each one fits.
Table: Olano vs Botpress
A developer-first platform you build and operate yourself, compared with a system built and operated for you.
Table: Olano vs Voiceflow
Designed conversations versus delegated operational work - and where Voiceflow's voice focus fits.
Table: Olano vs Grok Bot
Always-on bots sharing one cloud computer and every login, versus isolated agents on your customer channels.
Table: Olano Cortex vs Hermes self-learning
Two systems that learn from real work - one personal and local, one governed with approvals, undo and a budget.
Table: Olano vs OpenClaw
A self-hosted gateway you keep alive yourself, versus a managed deployment - and an honest look at what a VPS does and does not fix.
Table: Olano vs Meta Muse
Meta's personal agent on its own secure cloud VM, versus agents that belong to a business and answer its customers.
Table: Local vs cloud AI agents
A laptop is a session and a deployment is a service - what stops when the lid closes, and what renting a server does and does not fix.
Customer stories
How real businesses run agents day to day, in their own words.
An inventory-connected WhatsApp system
A motorcycle-parts retailer runs two WhatsApp agents on a custom inventory API - live prices, stock, and photos for customers, and stock updates from the staff's phones.
Tender intelligence for a marketing agency
Government tenders monitored twice a day, scored by fit, and delivered as email + PDF reports: "These reports are excellent. Thank you so much!"
Start from a workflow instead
Each use-case page shows the exact setup: what the agent does, which tools it connects to, and where the approvals sit.
WhatsApp AI customer service
Answer customer enquiries on WhatsApp around the clock, with approvals before replies go out.
AI inbox assistant
Triage email, draft responses, and keep follow-ups moving without losing threads.
AI market research agent
Research digests and reports on a schedule, grounded in sources your team can check.
AI social media agent
Drafts and content calendars that publish through connected tools only after you approve.
AI CRM sales assistant
Keep HubSpot, Salesforce, or Google Sheets updated and nudge deals that go quiet.
Missed-lead recovery
Catch the enquiries that arrive after hours and follow up before they go cold.
AI booking assistant
Take bookings, send reminders, and handle reschedules across your channels.
WhatsApp AI for clinics
Enquiries, bookings, reminders, and FAQs for clinics - never medical advice.
Not sure which workflow to start with?
Book a consultation. We map your highest-value workflow, quote it, and only build once you approve - live in days, not months.