🤖 Agents 101

What is an AI Agent?

An AI agent is an AI setup with a role, goal, rules and sometimes tools. Instead of only answering one prompt, an agent can follow a process, make decisions inside boundaries, use context, call tools, check outputs and move a task forward.

Browse Agents AI Workflow Guide Request a Workflow

Why agents matter

Agents are useful when the task needs more than a single answer. They can hold a role, follow instructions, use tools, make limited choices and hand back a result or recommendation.

Normal prompt

You ask once and the AI replies once.

Summarise these support tickets and tell me the top issues.
  • Good for quick answers.
  • Usually no persistent role or process.
  • May not check its own work.
  • Does not normally take controlled actions.

AI agent

The AI follows a defined role and process, sometimes with tools.

Support Triage Agent:
1. Read new tickets.
2. Classify issue type and urgency.
3. Flag missing information.
4. Draft suggested response.
5. Send high-risk cases for human approval.
  • Better for repeatable operational tasks.
  • Can use context, tools and approval rules.
  • Can follow a workflow, not just write an answer.
  • Can be connected to automation once tested.

What goes inside an agent?

A good agent is not just “be helpful”. It needs a job, scope, limits, inputs, outputs and checks. Otherwise it becomes a very confident intern with a flamethrower.

🎯

Role

What the agent is responsible for, such as triage, research, writing, review, planning or monitoring.

📌

Goal

The outcome the agent is trying to produce, not just the task it performs.

📥

Inputs

The data, files, prompts, tickets, emails, context or user instructions the agent needs.

📤

Outputs

The format it returns: summary, draft, decision brief, ticket update, table, JSON or action list.

🛡️

Boundaries

What the agent must not do, where it needs approval, and what it should escalate.

🔌

Tools

Optional access to systems, connectors, files, webhooks, calendars, docs, tickets or databases.

✅

Quality gate

A check before the agent’s output is trusted, sent, published, stored or actioned.

🧑‍⚖️

Human approval

The point where a person must approve, reject or edit the agent’s recommendation.

How an agent works

Agents work best when the route is clear. Give them a trigger, context, task, check and handoff.

1. Trigger A user request, form, ticket, file, schedule or webhook starts the agent.
2. Context The agent loads rules, background, examples and constraints.
3. Work step It drafts, classifies, researches, plans, reviews or extracts.
4. Tool step If allowed, it reads or writes to connected systems.
5. Quality gate The result is checked for risk, accuracy, format and completeness.
6. Handoff The agent returns output, asks for approval, or triggers the next workflow step.

A simple agent example

This is a basic agent pattern for support ticket triage. Simple, clear, testable.

Agent setup

Agent name:
Support Triage Agent

Role:
Classify incoming support tickets and prepare a suggested next action.

Goal:
Help the support team prioritise tickets without auto-sending customer responses.

Inputs:
- ticket title
- ticket body
- customer type
- product area
- previous ticket history, if available

Outputs:
- category
- urgency
- one-sentence summary
- missing information
- suggested response
- human approval needed: Yes / No

Agent rules

Rules:
- Do not reply directly to customers.
- Do not invent missing customer details.
- Mark billing, legal, security and outage issues as human approval required.
- If urgency is unclear, ask for more information.
- Keep summaries factual and short.

Quality gate:
Before final output, check:
1. Is the category supported by the ticket?
2. Is urgency justified?
3. Is any information missing?
4. Does this require human approval?

When should you use an agent?

Use an agent when a workflow needs a role, process, tools, decisions or handoffs. Use a prompt or skill when the task is simpler.

Use an agent when…

  • The task repeats and has a clear role.
  • The workflow has multiple steps or decision points.
  • The AI may need tools, connectors or files.
  • There are clear approval and escalation rules.
  • You want to move from manual work toward automation.

Use something simpler when…

  • You only need one answer.
  • A reusable skill is enough.
  • The task has no tools or handoffs.
  • The output is low-risk and manually reviewed.
  • The agent would add complexity without control.

Common agent types

Agents should be named by the job they do, not by vague magic. Useful agents have useful boundaries.

🔎

Research Agent

Collects information, extracts facts, flags unknowns and prepares summaries for review.

Browse research agents
🎫

Triage Agent

Classifies tickets, emails or requests and routes them based on urgency and rules.

Browse triage agents
🛡️

Review Agent

Checks output, code, risk, evidence, policy alignment or quality before handoff.

Browse review agents

Where to go next

Agents sit between skills and automation. They work best with context files, quality gates and clear human approval points.

🤖

Browse Agents

Find reusable agents for research, review, triage, planning, operations and automation workflows.

Open agents
📁

Context Files

Give agents the background, rules and constraints they need before they act.

Open context hub
✅

Quality Gates

Check agent output before it is sent, published, stored or automated.

Browse gates

Need an agent for your workflow?

Describe the role, task, systems, risks and approval points and get routed toward the right agent, skill, context file, quality gate or automation-ready workflow.

Request a Workflow