Skip to main content
A Singapore Government Agency Website How to identify
Official website links end with .gov.sg
Secure websites use HTTPS
Look for a lock () or https:// as an added precaution. Share sensitive information only on official, secure websites.
Government officials will never ask you to transfer money or disclose bank log-in details over a phone call.

Call the 24/7 ScamShield Helpline at 1799 if you are unsure if something is a scam.

Institute of Digital Government

AI Agents

Learning Pathway:

AI 301 Build

16 July 2026

What exactly are agents? How do they work? And can they really take work off your plate?

Illustration of an AI agent helping manage documents, emails, schedules, and business tasks through automated workflows and insights. Provide your feedback on BizChat

AI Agents

💡 One-line definition

An AI Agent is an AI that can plan, take actions, check its own work, and keep going, instead of just replying to one prompt at a time.

Think of the AI chatbots you already use (ChatGPT, Claude, etc.) as a vending machine: you put in a specific request, you get one specific thing back.

Picture of a vending machine

An AI Agent would be like your personal assistant, who doesn't wait for you to spell out every step and figures it out how to achieve the goal you've set on its own. Its able to utilise multiple tools/apps to accomplish the task at hand, check its own output and fixes these issues, rather than handing you the first draft.

A picture showing an efficient personal assistant

Where This Shows Up in Everyday Work

  • Chatbot: "Summarise this document." → You get a summary.

  • Agent: "Review these submissions against the checklist and flag any that need follow-up." → It reads each one, applies the criteria, compiles the flagged list, and tells you why each was flagged.

The difference isn't the AI getting "smarter" — it's the AI being given a process to follow rather than a single question to answer.

How an agent works

Agents run in a loop. The LLM makes sense of the task at hand and decides on the appropriate action to take, fetches the necessary Tools to take those actions, and relies on the Skills it has to execute the task well. This process is supported by the background information that is made available for the Agent via Connectors that are set-up.

Diagram illustrating the fours steps of an Agent Loop, Perceive, Plan, Act and Observe
An example of how an Agent could shorten the time it takes for routine work tasks

🔗 See it in action

Several AI Apply 201 guides already show the building blocks — a daily morning brief (opens in new tab), searching SharePoint with AI (opens in new tab), meeting minutes (opens in new tab). An agent is what happens when these steps are chained toward a specific goal.

Where agents fit VS Where agents don't

✓ A Good Fit

  • Multi-step tasks with a clear goal

  • Work that needs several tools or sources

  • Repeatable, well-scoped routines

  • First drafts and legwork you'll review

  • Tasks where mistakes are cheap to catch

✕ A Poor Fit

  • High-stakes or irreversible decisions

  • Vague goals with no clear "done"

  • Anything needing accountable human judgement

  • Sensitive data without proper access controls

  • Situations where an error can't be undone


Using agents responsibly in public service

An agent that can act carries more risk than a chatbot that only talks.

🛡️ Before you let an agent run

  • Keep a human accountable. The officer stays responsible for the outcome — not the tool.

  • Scope it narrowly. Give it the least access and the smallest job that gets it done.

  • Mind the connectors. An agent inherits whatever data and permissions you grant it.

  • Test before you trust. Watch it on low-stakes runs before turning the autonomy dial up.

  • Keep a record. Be able to see what it did, and why.