Agent Engineer vs AI Engineer vs ML Engineer
ZarifAgent engineer, AI engineer and ML engineer are three titles for work that overlaps but is judged differently. An ML engineer is judged on the model, an AI engineer on the product feature built around one, and an agent engineer on what happens when the model is allowed to act. Reading a posting for which of those it owns tells you more than the title does.
The postings compared here were opened on September 26, 2026. Roles close and ranges change, so open the link before relying on any detail.
The three jobs side by side
| Agent engineer | AI engineer | ML engineer | |
|---|---|---|---|
| Owns | The loop that lets a model choose actions: tools, context, state, guardrails and evaluation of outcomes | Product features built on foundation models: integration, retrieval, prompts and shipping to users | Models trained or tuned on the company's own data, and their path to production |
| Judged on, per postings | Agent quality and reliability in production, measured by evaluations | Shipped features and their use in production | Model performance on a business problem, such as fraud caught |
| Typical tools | Tool schemas, eval harnesses, tracing, orchestration, sandboxes | LLM APIs, web frameworks, retrieval, cloud platforms | Python data stack, PyTorch, statistical methods, feature pipelines |
| Usually sits in | An agents, agent platform or applied AI team in engineering | Applied AI or product engineering | A data or ML department close to the domain team |
| Main artifact | An agent that fails safely and an eval that proves it | A working AI feature | A model and the evidence that it beats the baseline |
These are defaults drawn from the postings below, not rules. The sections that follow source each column.
Agent engineer: owns what the model is allowed to do
Agent engineering postings talk about actions, tools and evaluation more than about models. The work splits into two layers: building agents for a product or customer, and building the runtime other agents run on.
On the application side, Harvey's Senior Software Engineer, Agents asks the engineer to "design environments and actions for agentic professional work, make model selection decisions, manage context windows, create optimal tools, and develop evals." Sierra's Software Engineer, Agent is responsible for "building, tuning, and evolving AI agents in production environments." Scale AI's Frontier Agents Engineer (Applied AI) names "agent guardrails, fallback strategies, tracing, monitoring, and evaluation pipelines."
On the runtime side, Sierra's Software Engineer, Agent Runtime sits in platform engineering and works on "the orchestration engine, runtime, and primitives that define how agents reason, take actions, and interact with users and systems." Anthropic's Staff+ Software Engineer, Claude Managed Agents designs "durable session and event storage, sandbox orchestration, streaming, scheduling, and multi-tenant isolation." That is distributed systems work, and the posting asks for at least eight years of it.
What these postings share is a concern with what goes wrong after the model decides. WRITER's Software engineer, agents puts it plainly: "Ensure reliability, monitoring, and observability for all agent components." Anthropic's own guidance, Building effective agents, describes agents as systems where the model directs its own process and tool use, as opposed to workflows that follow a fixed code path. An agent engineer is the person who decides where that line sits and makes the result safe to run.
The agent engineer career guide breaks the work into layers with a portfolio exercise for each.
AI engineer: ships features built on foundation models
AI engineer postings are broader. The engineer integrates models into a product and ships the result, and agents are one possible output among several.
Ramp's Applied AI Engineer asks the engineer to "ship full-stack AI projects end to end" and wants "a track record of working on full-stack AI projects, particularly those involving production use cases of LLMs." The same posting also covers "components for AI infrastructure, supporting production-level inference and fine-tuning" and internal tools for other engineers. The emphasis is breadth and shipping.
WRITER's AI engineer reports to the head of AI engineering and will "drive the development of intelligent agents and AI-powered features." It also asks for PyTorch, TensorFlow or JAX and "MLOps tools for managing the AI lifecycle." That is the overlap in one posting: agents are in scope, and so is model tooling that an agent engineer posting rarely mentions.
The practical difference from an agent engineer is depth on one question. An AI engineer is judged on whether the feature works for users. An agent engineer is judged on whether the actions the model takes are correct, permitted and recoverable. The AI engineer career guide covers the broader role with a retrieval and evaluation project.
ML engineer: owns the model
ML engineer postings start from data. The engineer trains, evaluates and deploys models for a specific problem, and the model itself is the product.
Ramp's Machine Learning Engineer sits in the data department and works with fraud engineering. The posting asks the engineer to "prototype and productionize machine learning models and rules-based systems to protect Ramp and its users from fraud." It wants "strong python experience (numpy, pandas, sklearn, pytorch etc.)" and "prior experience deploying Machine Learning models to production."
Compare that with the Applied AI Engineer posting from the same company. Both ship to production, and both need backend skills. The ML engineer works with statistics, feature data and a model the company owns. The applied AI engineer works with foundation models someone else trained. Neither posting mentions agent tool permissions or approval flows.
Where the titles blur
Three patterns show up often enough to watch for.
- The AI engineer who builds agents. WRITER's AI engineer posting includes "intelligent agents" in its scope. If most of the bullets are about tools, actions and evaluation, prepare for an agent engineering interview.
- The agent engineer who is a distributed systems engineer. Anthropic's Claude Managed Agents and Sierra's Agent Runtime postings are platform jobs. The agent knowledge matters, but the interview will test storage, isolation and failure modes at scale.
- The agent engineer who is customer-facing. Decagon's Senior Software Engineer, Agent Product will "integrate them at scale with large enterprise-grade customers." That shades toward forward deployed engineering, which the FDE vs solutions architect comparison covers.
The literal title Agent Engineer is still rare. On the boards checked for this piece, Scale AI's Frontier Agents Engineer and Elastic's Agentic AI Engineer came closest. Most employers post "Software Engineer" with an agents team name.
How to tell which one a posting really is
Read for these signals before the title. Most postings mix a little of each, so count which column wins.
| Signal in the posting | Leans agent engineer | Leans AI engineer | Leans ML engineer |
|---|---|---|---|
| Main verbs | Design tools, orchestrate, guard, evaluate agents | Integrate, ship, build features | Train, tune, model, analyze |
| Named skills | Tool calling, context management, eval harnesses, tracing | LLM APIs, full-stack frameworks, cloud platforms | Statistics, sklearn, PyTorch, feature engineering |
| Data it works with | Traces, tool results, task outcomes | Prompts, documents, user feedback | Labeled datasets, features, model metrics |
| Failure it worries about | A wrong or unauthorized action, a stuck loop | A bad answer or a slow feature | Drift, false positives, a model that stops generalizing |
| Team name | Agents, agent platform, agent runtime | Applied AI, AI product | Data, ML platform, a domain team like fraud |
When the posting leaves the answer open, ask in the first call:
- Does the model in this product take actions that change records, or only return text?
- Who owns the evaluation set, and how is agent quality measured today?
- Does the team train or fine-tune models, or only use hosted ones?
- What broke in production last quarter, and who fixed it?
Posted ranges, read carefully
These ranges come from the postings linked above, opened September 26, 2026. They are examples of how specific employers post pay, not market averages or comparable levels.
| Employer and role | Posted range | What the posting says it covers |
|---|---|---|
| Ramp, Applied AI Engineer | $204.4K – $352K | Compensation field, plus equity. Not labeled base salary |
| Ramp, Machine Learning Engineer | Estimated Base Salary $200K – $330K | Base salary, plus equity |
| WRITER, AI engineer | SF & NYC Base Compensation $220K – $259K | Base compensation in SF and NYC, plus equity. Other US locations: $198K – $233K |
| WRITER, Software engineer, agents | SF & NYC Base compensation $146K – $240K | Base compensation in SF and NYC, open to multiple levels, plus equity. Other US locations: $132K – $216K |
| Sierra, Software Engineer, Agent Runtime | $230K – $390K | Compensation field, plus equity. Not labeled base salary |
| Anthropic, Staff+ Software Engineer, Claude Managed Agents | Annual Salary: $405,000 - $485,000 USD | Annual salary for a staff-plus role requiring at least eight years of experience |
The Ramp pair is the most direct comparison, because the same company posts both roles in the same city. The two WRITER ranges differ in level as well as role: the agents posting spans several levels, so its lower bound is not a like-for-like comparison with the AI engineer range. The Anthropic row reflects seniority more than the agent specialty. More agent engineering ranges are on the Agent Engineer career path.
Which one fits your background
If you come from backend or distributed systems work, agent engineering is the closest step, especially the runtime side. The gap is evaluation: learning to measure a task whose output varies run to run.
If you come from product or full-stack engineering, AI engineering is the natural entry, and agent work is a specialization inside it. Start with one feature, then add one tool the model can call.
If you come from data science or statistics, ML engineering uses what you have. Moving to agent work means learning to run a service: authentication, retries and writes that are safe to repeat.
The how to become an agent engineer guide maps each background to lessons in the site's agent course.
Next steps
- Follow the Agent Engineer career path: reading order, posted pay and practice prompts.
- Read the agent engineer career guide for the layers of the work.
- Prepare with the agent engineer interview guide.
- Compare the broader role in the AI engineer career guide.