Agentic AI products
Agents don't just answer — they act. This module shows product managers how to scope, design, govern and measure AI agents.
Learning objectives
- Distinguish workflows from agents and choose the simpler option
- Scope an agent's goal, tools, permissions and stop conditions
- Set autonomy per tool by risk
- Define metrics for agent performance
3.1 From features to agents
An AI agent uses a model, tools and a loop to pursue a goal: it decides the next step, calls a tool (search, a database, an email API), looks at the result and continues until it's done or stuck. Anthropic's guidance on building effective agents distinguishes workflows (the steps are fixed in code) from agents (the model chooses the steps), and recommends the simplest option that works.
OpenAI's practical guide to building agents suggests agents fit best where traditional automation struggles: complex judgement calls, rules that are hard to maintain, and heavy use of unstructured information. If a fixed workflow does the job, ship that.
3.2 Scoping an agent like a job description
| Goal | What "done" means, in measurable terms |
| Tools | Exactly which systems it may read and change |
| Permissions | Spending limits, which records, which customers |
| Stop conditions | Maximum steps, time or cost; when to give up |
| Escalation | When and how to hand over to a person, with context |
Start with one agent and a small tool set; add more agents only when one becomes unmanageable.
3.3 Risk-rated tools and autonomy
| Tool risk | Examples | Typical oversight |
|---|---|---|
| Read | Search knowledge base, look up an order | None needed beyond access control |
| Reversible write | Draft a reply, tag a ticket, create a calendar hold | Act and notify; easy undo |
| Irreversible or external | Send an email to a customer, delete records | Human approval until quality is proven |
| Financial or high-stakes | Issue refunds, place orders, change contracts | Approval above a limit; hard caps always |
Autonomy is granted per tool, not for the whole agent. An agent can read freely, draft freely and still need sign-off to send.
3.4 Designing trust for agents
- Show the plan before acting on anything significant, and let users edit it
- Show progress — which step it's on and what it has found
- Checkpoints before risky actions, with a clear summary of what will happen
- An activity log of every action taken, with undo where possible
- A graceful handover to a human that includes everything the agent learned
3.5 Measuring agents
| Metric | Question it answers |
|---|---|
| Task success rate | How often does it fully achieve the goal? |
| Intervention rate | How often must a person step in or correct it? |
| Cost and steps per task | Is it efficient, or wandering? |
| Time to completion | Is it faster than the human baseline? |
| Error severity | When it fails, how bad is it — wasted time, or real harm? |
| Tool-use accuracy | Does it choose the right tool with the right inputs? |
3.6 Your product as a tool for other agents
Agents increasingly act on users' behalf across many products. Standards like the Model Context Protocol (MCP) let an agent connect to your product's data and actions. Ask: should our product offer an MCP server or API that other agents can use — and with what permissions? The technical details are in AI Agents & Agentic Workflows, and the risks in AI Security & Safety.
Scope an agent for a task in your organisation: goal, tools with their risk rating, permissions, stop conditions and escalation. Then decide which tools need human approval at launch and what evidence would let you remove that approval.
See a worked example
Example — invoice-chasing agent: Goal: overdue invoices paid or escalated within 14 days. Tools: read ledger (read), draft reminder (reversible), send email (external — approval at launch), add note to account (reversible). Stop: 3 reminders or customer reply. Escalate: disputes, amounts over £10,000, any angry reply. Remove email approval after 500 approved drafts with under 2% edits and zero wrong-customer errors.
Key takeaways
- Use a fixed workflow when it does the job.
- Scope agents like a job: goal, tools, permissions, stops, escalation.
- Grant autonomy per tool, by risk.
- Show the plan, progress, checkpoints and an activity log.
- Measure task success, interventions, cost per task and error severity.