What 'Build AI Agents' Means in AI Job Postings (160 Companies)
By Saurav Sharma · November 18, 2026
Job postings say "build AI agents" without saying what the work is. We searched every posting in our dataset that mentions agents and used keywords to sort the sentences into eight kinds of work. Here, an agent is software that uses a model to choose and carry out steps toward a goal. This guide shows how often each kind appears, quotes short pieces of real postings, and gives you what to explain in an interview and a weekend project for each one.
What the data is
The numbers come from aijobprep's study of 3,647 AI-engineering job postings at 198 companies, collected August 2026. Agents are one of the top topics: 1,348 postings at 160 companies match at least one of our agent search terms. This is a snapshot from one month, so check current listings before you rely on a number.
How we counted: we searched for the words agent, agents, agentic, "tool use" and "agent harness" after removing standard text that repeats across a company's postings. We then checked each matching sentence for task keywords. A sentence with keywords for several task types counts in each of those rows, so the rows overlap. The counts show what postings mention. A mention does not prove a company requires or tests that skill.
The eight kinds of agent work
| Task type | Postings | Companies |
|---|---|---|
| Orchestration and multi-step workflows | 730 | 127 |
| Tool calling and integrations | 640 | 126 |
| Customer-facing deployment | 599 | 109 |
| Evaluating agent behavior | 428 | 95 |
| Reliability, guardrails and human approval | 398 | 100 |
| Agent platforms for other teams | 363 | 99 |
| Observability and tracing | 202 | 69 |
| Cost and latency | 128 | 50 |
In total, 1,137 postings at 147 companies have at least one agent sentence that fits one of these types. Orchestration and tool calling lead. Tracing agent runs and managing cost or response time are the two least common types.
Do this now: open three postings for the role you want, highlight sentences about agents or tool use, and mark the task types they describe. If a sentence names no task, write down what you would ask the hiring team.
The vague mentions
Not every mention names work. Of the 1,348 postings that mention agents, 211 do not fit any of the eight types. Separately, 153 postings at 56 companies have at least one sentence with soft wording such as "familiarity", "exposure" or "experience with" and no task named. When you see "agentic mindset" or "familiarity with agents", the posting does not tell you the work. Ask the recruiter or hiring manager which of the eight types the team does today, and prepare for that one.
Each task type: what to explain and what to build
The quotes are short fragments from postings. The interview points and projects are our recommendations based on the task types. They do not describe any company's actual interview.
Tool calling and integrations (640 postings at 126 companies)
Postings mention, for example: "tools/MCP integrations, retrieval, and evaluation" (Databricks); "Proficient in tool use, RAG pipelines" (Nvidia); "tool calling, and context management" (Roblox).
Be ready to explain: How a model picks a tool, how you describe the tool so it picks well, what happens when the call fails or returns junk, and how you check the inputs the model passes to a tool before the tool runs. Know what the Model Context Protocol (MCP) is: a standard way to connect tools to a model. Know what retrieval-augmented generation (RAG) is: looking up relevant documents and giving them to the model before it answers.
Weekend project: Build an agent with three tools (search, a calculator, and a file reader) behind a small MCP server. Log every call and add a test where a tool times out. Write down what the agent did next.
Orchestration and multi-step workflows (730 postings at 127 companies)
Postings mention, for example: "agents that plan and execute multi-step tasks" (Cohere); "persist context across long-running tasks" (OpenAI); "multi-agent orchestration patterns" (GitLab).
Be ready to explain: When one model call is enough and when a loop or a fixed workflow is better. How you keep state between steps, how you stop a loop, and how you resume after a crash. Be ready to explain when one agent is better than several.
Weekend project: Build the same task twice: a fixed three-step workflow and a free-running agent loop. Run each 20 times. Compare success rate, number of model calls, and cost in a short table.
Customer-facing deployment (599 postings at 109 companies)
Postings mention, for example: "collaborate directly with customers" (Decagon); "alongside customer teams" (Nebius); "adoption of GitLab Duo Agent Platform with strategic customers" (GitLab).
Be ready to explain: How you turn a vague customer request into a scoped agent with a clear success measure. What you ask in the first call, how you handle data access, and how you explain a failure to a non-technical buyer.
Weekend project: Pick a fake customer with a messy request, such as sorting support emails. Write a one-page scope with success measures, then build the smallest version and a demo script that shows one failure on purpose.
Evaluating agent behavior (428 postings at 95 companies)
Postings mention, for example: "building and evaluating complex agentic or multi-step LLM workflows" (Cresta); "benchmarks and regression tests for ... agent performance" (Datadog); "evaluation platforms for agent behavior validation" (Airbnb).
Be ready to explain: How you build a test set from real cases, how you score a multi-step run (final answer, each step, or both), when you would trust a large language model (LLM) to grade a result, and how you check that a change did not break something that used to work.
Weekend project: Write 30 test cases for your tool-calling agent. Give each run two scores: did it produce the right answer, and did it call the right tools in a sensible order? Change the prompt, rerun, and show which cases got worse.
Reliability, guardrails and human approval (398 postings at 100 companies)
Postings mention, for example: "fast, safe iteration on agent behavior" (Decagon); "the right guardrails, controls, and visibility" (Glean); "human-in-the-loop checkpoints" (Nvidia).
Be ready to explain: Which actions need a person to approve, how you limit what an agent can touch (permissions, budgets, step limits), and how you handle prompt injection, where text in a tool result tries to give the agent new orders.
Weekend project: Add an approval step to any agent that sends email or writes a file. Add a step limit and a spend limit. Then feed it a web page that contains a hidden instruction and record what your guardrails did.
Agent platforms for other teams (363 postings at 99 companies)
Postings mention, for example: "the cloud agent platform that powers them" (Figma); "code-execution sandboxes, eval harnesses" (Anthropic); "containerized agent platform" (Docker).
Be ready to explain: What a shared platform gives other teams: a common way to define tools, run agents in a sandbox, set permissions, and see results. Be ready to explain how a team adds its first agent, releases an update, and keeps that agent away from another team's data.
Weekend project: Build a small runner that loads an agent from a config file, runs it in a Docker container with no network except one allowed host, and writes a result file. Add a second agent config to show a team could reuse it.
Observability and tracing (202 postings at 69 companies)
Postings mention, for example: "Automate monitoring of agent behavior using observability tools" (Adobe); "agent observability" (GitLab); "evaluation, observability, and production reliability" (Navan).
Be ready to explain: What you record per run: prompts, tool calls, results, token counts, and timing, tied together by one run ID. Explain how you would find the one failing step in a 15-step run. A trace is the step-by-step record of one run.
Weekend project: Add tracing (a step-by-step log of prompts, tool calls, and results) to your agent with OpenTelemetry or a tool such as LangSmith. Break the agent on purpose, then find the failing step from the trace alone and write a three-line incident note.
Cost and latency (128 postings at 50 companies)
Postings mention, for example: "managing latency, context windows, and model choice" (Abridge); "cost spikes, permission edge cases in MCP servers" (Drata); "latency, cost, safety, security, and governance" (OpenAI).
Be ready to explain: Where agent cost comes from (repeated context, long loops, large models used for easy steps), and what you would change first: caching, a smaller model for simple steps, shorter context, or a step cap.
Weekend project: Measure tokens, time, and dollars per run for your agent. Make two changes, such as a smaller model for one step and a trimmed context. Report the before and after numbers and the effect on eval scores.
A simple way to use this
- Start with an agent that can call one tool, such as a calculator or a search API. You can reuse it for the testing, tracing, and cost projects.
- Add evals and traces to it, since those two types let you debug everything else.
- Add an approval step and a cost report.
- For each project, write down one failure, how you found the cause, and what you changed. Use it to practice explaining your work.
If you are aiming for a customer-facing role, practice the customer-facing project first. See the forward-deployed engineer questions. For product roles, the eval and guardrail sections matter most: AI product manager questions. Engineers can start with the AI engineer questions and ML engineer questions. For a longer view of skills, read which AI engineer skills to stop learning.
Practice explaining agent work out loud. Start with the AI engineer interview questions.