Loading...
System Design and AI Engineering, by Roni Das. 980 interactive lessons.
Start LearningAn agentic AI engineer builds AI agents: software where a language model decides which tools to call, and in what order, to finish a task. Anthropic describes agents as systems where models "dynamically direct their own processes and tool usage". The job is making that loop safe and reliable in production: tool contracts, memory and context, permissions, evals that measure whole tasks, and cost.
Preparing for an agent engineering role? The AI Engineering course has a full chapter on agents in production, plus evals and security, with measured experiments, not just theory.
Most of the work is not the model. It is the harness around it: which tools exist, how each tool is described, what the model sees at each step, what happens when a tool fails, and when to stop. Anthropic calls the tool side the agent-computer interface, and advises writing tool descriptions like "a great docstring for a junior developer on your team".
The second half of the job is proof. A demo that works once is easy. An agent that finishes the task nine times out of ten, and fails safely the tenth time, needs evals over many realistic tasks, traces of every step, and someone reading the failures.
The third part is safety. An agent with tools can send emails, change records and spend money. Permissions, sandboxes and defences against prompt injection are part of the design from day one, not added at the end.
“The autonomous nature of agents means higher costs, and the potential for compounding errors.”
“Engineer agent harnesses that connect powerful models with the environments and tools required to perform economically valuable work for users.”
“This team owns the improvement loop: define what successful work looks like, build evaluations that measure it, understand where agents fail, and ship changes that improve outcomes.”
“We're looking for applied AI engineers to help bring Codex agents from impressive demos to dependable tools.”
Quotes copied from the pages linked, read on 23 October 2026.
How a model chooses the next step, the ReAct and plan-and-execute patterns, and why errors add up as tasks get longer.
A tool call is input the model made up. Check it against a schema and the user's permissions in code before it runs. Anthropic advises "thorough tool documentation and testing".
The Model Context Protocol is a standard way to plug tools and data into a model. Know what it gives you, what it costs, and its security risks, such as a tool description that hides instructions.
What the model sees at each step decides what it can do. Long contexts get worse, so you summarise, select and drop old steps on purpose.
Score the task, not one answer. Run each task several times: an agent that passes 95% of single tries can still fail often over many steps. Read the failing traces before you build metrics.
Prompt injection, sandboxes, whose authority an agent acts with, and a person approving risky actions. Anthropic recommends "extensive testing in sandboxed environments, along with the appropriate guardrails".
For a software engineer who can already call an LLM API. About 8 to 10 hours a week, a rough guide and not a promise. Free lessons are marked; the rest are in the AI Engineering course.
Days 1 to 30
Before production problems, you need a loop you understand: the model, the tools and the format of every call.
Prove it: Build an agent with two tools, log every step, and count how often it finishes 20 tasks.
Days 31 to 60
Real tasks take minutes, call many tools and fail halfway. This month is about runs that survive that.
Prove it: Make your agent resume after a crash in the middle of a task, without repeating a tool call that already ran.
Days 61 to 90
This is what separates a demo from a job: proving the agent works, and proving it cannot be turned against its user.
Prove it: Hide an instruction in a document your agent reads, and show it does not follow it. Then write 50 eval tasks and report the score before and after one change.
The postings we read do not publish their interview steps, so we will not invent a loop. Here is what each says it looks for. Expect a coding round, an agent design round like the worked answer below, and questions about evals and failures you have debugged.
Building agents for real work
Perplexity asks you to "engineer agent harnesses that connect powerful models with the environments and tools required".
Evals and failure analysis
OpenAI's Codex role says you will "develop and run evals to measure agent performance, regressions, failure modes, and edge cases".
Reliability in production
OpenAI's Codex role: "Analyze failures in production and systematically improve robustness and reliability."
Security by design
Perplexity lists "developing permission architectures, payload classifiers, and other methods to implement secure-by-design agentic capabilities".
Here is the shape of a good answer to a typical agent design prompt. The point is the order: decide if it needs an agent at all, then tools, safety, evals.
Design a support agent that can look up orders and issue refunds.
One worked answer is not enough for the loop.
Each question has why it is asked, a full answer, what the interviewer listens for, the follow-ups, the common mistakes and a diagram (153 diagrams in all). A PDF you keep, junior to staff level.
India: UPI, cards or netbanking through Razorpay. Elsewhere: card or PayPal. Download at once from My packs.
Each range is the one the company's own job board showed on 23 October 2026, copied word for word. All are US postings and all say they also offer equity, which is not in the range.
| Company and title | Where | Published range |
|---|---|---|
| OpenAI, Applied AI Engineer, Codex Core Agent | San Francisco | $230K to $385K |
| OpenAI, Software Engineer, Agent Infrastructure | San Francisco | $230K to $385K |
| Perplexity, AI Software Engineer, Agents | San Francisco | $220K to $405K |
| Perplexity, Software Engineer, Enterprise Agent Evals and Efficiency | San Francisco | $220K to $405K |
The AI Engineering course has 210 lessons, and 10 of them are free to read first, with no card needed. The agents chapter covers the loop, tool contracts, durable execution, memory, MCP and multi-agent systems, and the evals and security chapters cover the rest of this roadmap. One payment, lifetime access.
Free 15-min Interview Readiness Check
A short call. I tell you the 2 or 3 things to fix first for your target role. It costs nothing, and it is booked on Topmate.
Also useful: AI engineer roadmap, Agentic AI vs generative AI, MCP interview questions, LangChain vs LangGraph and Forward deployed engineer guide.
