Agentic AI engineer: skills, responsibilities and portfolio projects

Anthropic’s engineering guidance distinguishes predefined workflows from systems in which a model dynamically directs its own process and tool use. This is a useful design distinction, even though job advertisements often use “agent” more loosely.

For a predictable form-processing task, a fixed sequence may be easier to test. An agent may be useful when the next action depends on intermediate findings. The decision should follow the task’s variability, cost and consequences, rather than a desire to use a fashionable architecture.

Build on ordinary software engineering. An agent that can call a tool still needs authentication, durable state, predictable interfaces and a safe response when a dependency fails. Framework familiarity helps only when you understand the behaviour underneath.

Use a fictional support queue and a small policy library. Allow the agent to classify an issue, retrieve relevant policy and draft a suggested response. Keep refunds, account changes and outgoing messages behind a human approval step.

Give the system an explicit list of allowed tools. Include a case where a tool times out, a record is missing, a duplicate request arrives and the supporting document contains misleading instructions. The expected result can be escalation rather than task completion.

Your report should show the starting state, the sequence of tool calls, the ending state and whether approval was required. Measure cost and duration over several runs because repeated agent executions can take different paths. These are proposed practice exercises, not a benchmark of any commercial system.

Anthropic’s January 2026 evaluation article explains why agent evaluation needs attention to tasks, grading and execution traces. Use that as background, then define success for your own application.

For the support example, a fluent answer is insufficient if it references the wrong customer or bypasses approval. Score classification, evidence use, allowed actions and the final state separately. Include ordinary requests as well as adverse cases, and keep a held-out set that you do not use for prompt tuning.

Review disagreements between automated grades and human judgment. A grading model can be wrong too. Record the version of prompts, tools and configuration with each result so a reviewer can compare changes.

Answer with examples from your own project. If you have only built a prototype, say so and identify the additional work needed before production use. Clarity about limits is part of engineering competence.

A backend engineer may enter through tool integration and state management. A platform engineer may focus on deployment, observability and runtime controls. An application security engineer may focus on permission boundaries and untrusted inputs. These are useful adjacent strengths, not separate guaranteed hiring tracks.

Search for responsibilities involving tool-using systems, workflow automation and AI application infrastructure as well as the exact phrase “agentic AI engineer”. A role’s daily work is more informative than its title.