AI agent engineer
From a single model call to agents that use tools, work in multiple steps, and can be measured, secured and operated. Builds on the AI engineer roadmap.
Sign in to track how many of the 5 roadmap tasks you have done.
For people who have already built a first application on top of a language model and want to build systems that act: they call tools, work in multiple steps, and can be measured and secured. The roadmap goes from deciding whether you need an agent at all, through tools, MCP, memory and orchestration, to evaluation, security and operations. By the end you will be able to design, build and defend an agent that a company trusts with real work.
Steps
Step 1: AI engineering basics
An agent is an application built on a language model, so you first need the basics from the AI engineer roadmap: Python, APIs, how LLMs work, prompt and context engineering, embeddings and RAG. You are done when you have built a simple application on top of a model with RAG and can estimate its cost per query.
Topics on the portal
Resources
Step 2: When to use an agent and when a plain workflow
external resourcesMost tasks are solved by a fixed procedure (a chain of steps, routing, parallelisation). An agent with a loop and free choice of tools is more expensive and less predictable. Learn to recognise which pattern a task needs. You are done when, for three real assignments, you design the architecture and justify why a workflow is enough and where an agent is unavoidable.
Resources
Step 3: Tools and function calling
external resourcesAn agent acts through tools: functions with a clear description and an input schema, which the model calls and whose results it reads. A good tool description is half the success, a bad one leads to nonsensical calls. You are done when an agent with four tools (search, calculation, reading a file, writing) reliably completes ten test assignments and every tool call has a validated input.
Resources
Step 4: Model Context Protocol and integrations
external resourcesMCP is an open standard through which agents access data and services (files, databases, company systems) without a custom connector for every tool. You are done when you write your own MCP server with two tools and one data source and connect it to two different clients.
Resources
Step 5: Memory, state and long tasks
external resourcesThe context window is limited and expensive. An agent needs working memory (what it is doing right now), long-term memory (what it has learned about the user and the project) and the ability to pause and resume a task. You are done when your agent handles a twenty-step task without exceeding the context and, after a restart, continues where it left off.
Resources
Step 6: Orchestrating multiple agents
external resourcesMore complex systems split the work between specialised agents: a planner, executors, a reviewer. The key ingredients are clear interfaces between them, handing over context, and limits (number of steps, token budget, time cap). You are done when you build a system with an orchestrator and two worker agents that has hard limits and stops gracefully when they are exceeded.
Resources
Step 7: Evaluation and observability
An agent you do not measure fails silently. You need a set of test tasks with expected results, automatic grading (including a model as the judge), tracking of every step (trace), cost and latency, and regression tests on every change of prompt or model. You are done when you have an evaluation set with at least fifty cases and a dashboard showing success rate, cost and the most common failures.
Topics on the portal
Resources
Step 8: Agent security
An agent with tools is an attack surface: prompt injection through documents and web pages, data leaks, tool abuse, infinite loops. The principle of least privilege, human confirmation of dangerous actions, an isolated environment and limits are mandatory. You are done when your agent passes a set of prompt injection attacks without performing an unauthorised action.
Topics on the portal
Resources
Step 9: Deployment, costs and operations
optionalexternal resourcesIn production, reliability and price decide: choosing a model per task, caching, streaming, limits and retries on errors, prompt versioning, gradual rollout, and a plan for what happens when the model provider is unavailable. You are done when you have an agent deployed with monitoring, a daily budget and a fallback model.
Resources