Career path
Agent Engineer
An agent engineer builds the software that lets a model take actions. That means the tools it can call, the limits on what each tool may do, saved progress that survives a crash, and the tests that show the agent did the right thing. Employers rarely use the exact title, so read what a posting says you'd own. Use this path to understand the work, build proof with the site's agent course and prepare for the interview.
Read these 10 guides in order
- Agent Engineer: A Practical Career and Portfolio Guide
See the parts of the job, and try the exercise where an agent has to recover after something fails.
- Agent Engineer vs AI Engineer vs ML Engineer
Tell an agent role from an AI or ML engineering role before you apply.
- How to Become an Agent Engineer
Pick the lessons that fill your gaps and the three portfolio pieces to build.
- Run a Bounded Tool-Calling Loop in Python
Keep what the model asks to do separate from what your app actually allows.
- Build an MCP Server and Client for a Ticket-Label Tool
Offer a tool over MCP, a standard way to connect agents to tools, and test what happens when the inputs or the connection go wrong.
- Human Approval for Agents: Store and Check the Exact Decision
Tie each approval to the exact change, and reject requests that are out of date or repeated.
- Evaluate a Ticket Agent: Dataset, Scoring Code and Failure Report
Score an agent without dropping the runs that failed.
- Agent Design Patterns: Run Four Control Flows and Their Failures
Decide when fixed steps beat letting the model choose its next tool, and see how each one breaks.
- Build a Ticket Workflow with Signed Delivery and Approval
Wire incoming tickets, document lookup and approval together, then check that a restart or a repeat doesn't break anything.
- Agent Engineer Interview Guide
Prepare for the agent design, evaluation, debugging and coding rounds.
Pay in employer job postings
Checked September 26, 2026. These are yearly ranges from 6 real job postings. They aren't market averages, the jobs aren't all at the same level, and no one is promised these numbers. Open the posting for the current range and the full package.
| Employer and role | Location | Posted range | What the range covers |
|---|---|---|---|
| Scale AI Frontier Agents Engineer (Applied AI) | San Francisco / New York | $180,000 - $225,000 USD | Base salary that may span several career levels; equity and benefits are separate parts of the package. |
| Elastic Agentic AI Engineer | United States (distributed) | $94,300—$149,200 USD | Typical starting base salary with no variable compensation; select locations including the San Francisco Bay Area and New York City metro use $113,300—$179,200 USD. Eligible for the stock program. |
| Sierra Software Engineer, Agent | San Francisco / Atlanta / New York (on-site) | $180K – $390K | The posting's compensation field, plus equity. It does not say whether the range is base salary. |
| Decagon Senior Software Engineer, Agent Product | San Francisco (on-site) | $200K – $400K | Base salary, plus equity. |
| Harvey Senior Software Engineer, Agents | San Francisco (hybrid) | $193,400 - $290,000 USD | Compensation range for hiring across mid, senior and staff levels; equity and a bonus are offered in addition. It does not say whether the range is base salary. |
| WRITER Software engineer, agents | New York City / San Francisco / Seattle (hybrid) | $146K – $240K | Base compensation in San Francisco and New York City, open to multiple levels; other US locations are $132K – $216K. Equity may be offered separately. |
Interview practice
Practice questions written for this site, not leaked questions or any employer’s real interview. Open one to see what a strong answer covers.
An agent's tool call changed a record, but the worker crashed before saving the result. What happens next?
Point to the exact moment things could have gone wrong. Use a stable ID, or idempotency key, to check whether the change already happened, and explain why retrying without checking can make the change twice.
The model passes a valid-looking ID for a record the user isn't allowed to see. Where is that stopped?
Follow the user's identity from the request to the tool, have the app check who owns the record on every read and write, and explain why an instruction in the prompt is not access control.
Design the tools for an agent that updates customer accounts.
Prefer small tools with clear inputs, name the actions that need a person to approve them, tie each approval to the exact change, and make every tool's errors visible.
The test score improved, but complaints about the agent went up. What do you check?
Check whether the test set looks like real tasks, check what actually changed in the system rather than the agent's own summary, test whether the grader accepts confident wrong answers, and add the cases it missed.
An agent keeps repeating the same failing tool call. How do you fix it?
Find out why the rule that should stop it didn't fire, make the tool's error say what went wrong, cap the number of steps, show a clear failure, and add a test that catches the loop.
Which parts of this task should not be an agent?
Keep the steps that never change in ordinary code. Name the one decision that needs a model and the evidence that the model does it better, and compare against a fixed workflow.
Other AI engineering roles
12-guide path
Forward Deployed Engineer
Work with a customer, find out what is actually stuck, and build the software that makes their workflow run.
10-guide path
AI Engineer
Build useful software on top of AI models, then find out where it gets things wrong and fix it.
5-guide path
GTM Engineer
Fix the data and automate the handoffs a sales or marketing team depends on.
