This AI agents training shows how to design agents with tool access so that they stay limited, traceable and controllable, and it is aimed at developers, architects and technical product owners. You design permissions on the principle of least privilege, build in approval steps, recognise prompt injection and evaluate agent runs with test cases. You need programming experience and first hands-on work with language models. Scope and platform are agreed when you enquire.
What you’ll take away.
- Tell workflows from agents and decide when a simple flow is enough
- Design tool access on least privilege with separate read and write rights
- Define human approval steps for risky actions
- Recognise and mitigate prompt injection and excessive agency using the OWASP list
- Evaluate agent runs with test cases, logs and stop criteria
Your training in detail
- Who is it for?
- Software developers, architects, security leads and technical product owners who plan or build agents.
- Prerequisites
- Programming experience, preferably Python, and first experience with language model APIs and JSON.
- Format & planning
- Suggested format: 2 days live online or on site, split into modules. Scope depends on prior knowledge.
Agenda: from understanding to application
Our independently developed agenda proposal. Focus areas, lab environment and scope are agreed before booking.
Patterns from Anthropic's guide Building effective agents: prompt chaining, routing, parallelisation, orchestrator and evaluator.
Tool definitions with function calling, JSON schemas and integration via the Model Context Protocol.
Permissions and isolation: least privilege, separate credentials, sandboxes and output limits.
Approvals and control with human-in-the-loop, confirmation dialogs, budget and step limits.
Attack surfaces: prompt injection, excessive agency and data leakage per the OWASP LLM Top 10.
Evaluation with test cases, logging of intermediate steps, error rates and regression tests.
Suggested exercise: You build a small ticket agent with two read tools and one write tool, add an approval before writing and test it with a planted injection instruction.
Exams, licences & credentials
There is no uniform external exam for building agents safely. Attending alone awards no vendor certificate. Exam fees and licences are only included if the offer confirms them.
Vendor references & review date
Sources checked on 2026-11-03. These sources describe the technology or external exam requirements. They do not imply a partnership with NextNowa.
Your learning content
These objectives describe the training topic. The specific syllabus, duration, dates, trainer and price are agreed before booking.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in conversation. An agent can additionally call tools, such as querying databases, drafting emails or changing tickets, and plan several steps itself. That raises the benefit but also the risk, which is why permissions, approvals and testing are central.
How do I protect an agent against prompt injection?
There is no complete protection. You limit the damage: least privilege, separate trust zones for third-party content, approval before risky actions and logging. You also test deliberately with attack examples. The OWASP list serves as a checklist that we apply practically in the training.
When is an agent worthwhile and when is a fixed workflow enough?
When the steps are known in advance, a fixed workflow is usually simpler, cheaper and easier to verify. An agent pays off for open-ended tasks whose path only emerges during processing. The advice is to start with the simplest solution and add autonomy only when needed.
Does Designing safer AI agents include a certificate?
There is no uniform external exam for building agents safely. Attending alone awards no vendor certificate. Exam fees and licences are only included if the offer confirms them.