11 min read
Since the beginning of October, we’ve analyzed 1,376 open AI engineering roles across 159 AI companies.
One trend stands out immediately: forward deployed engineering has become the largest category in the market.
We found 451 open forward deployed roles across 77 companies—roughly one in three AI engineering jobs in our dataset, and comfortably ahead of the next-largest category.
Our dataset isn’t exhaustive, but it spans a broad range of AI companies and industries, giving us a useful snapshot of where engineering demand is concentrated right now.
So what does the AI engineering job market actually look like, and where is demand heading?
Let’s dive in.
Forward deployed engineering has 451 open roles (33%), and Agent engineer comes second with 258 (19%).
The next four families are AI platform, dev tools / infra, research engineer and applied AI, with between 51 and 94 roles each.

The title varies from company to company: Forward Deployed Engineer, Deployed Engineer, Applied AI Engineer, Deployment Strategist, Agent Deployment Engineer. The work is the same. You sit with a customer, build the agent on their data and systems, and stay until it runs in production.
The role started at Palantir.
According to The Pragmatic Engineer, Palantir created it in the early 2010s, and until about 2016 it had more forward deployed engineers than "normal" software engineers.
The biggest hirers for this role are OpenAI with 66, Wonderful with 49, LangChain with 28, Anthropic with 27 and Sierra with 26.

This is the most common AI engineering role, and it has the lowest median posted pay of the 11 role families with at least 20 US pay ranges.
679 roles post a US yearly range.
That's 49% of the total, and many of them are there because state law requires it.
California's SB 1162 has required employers with 15 or more employees to put a pay range in job postings since January 2023. New York State has done the same for employers with four or more employees since September 2023, and Washington requires a wage scale or salary range on postings.
The figures below are the median of the midpoint of each posted US yearly range. They leave out equity and bonus, which at these companies can be a large share of total pay.

The two big labs raise the top of the scale.
Anthropic's median midpoint is $445k and OpenAI's is $334k, while every other company together sits at $245k. Leave both labs out and the gap is still there: agent engineers at $267k, forward deployed engineers at $225k.

33% of forward deployed ads mention travel.
Across every other AI engineering role, it's 7%. 93 ads say how much travel. The median maximum is 40%, and the most common answer, in 39 ads, is up to 50%. (When an ad gives a range like "25-50%", we count the higher figure.)

The ads put it plainly:
Does it fit you? If you like putting things into production in front of a user, can handle a messy client codebase, and can be away from home one or two weeks a month, this is the largest pool of open roles you'll find. If you want deep, long-running ownership of one system, or travel doesn't work for you right now, agent engineer is the next-largest family and its median posted pay is higher. Our walkthrough of the forward deployed and applied AI interview loop covers each round if you decide to go for it.
17 of 1,376 roles require a PhD, and 14 more prefer one. 907 ads (66%) don't mention a degree at all.
Years of experience is where the bar sits.
789 ads state a minimum. The median is 5 years, 68% ask for 5 or more, and 72 (9%) accept two years or fewer. Only 25 roles have intern, new grad, early career or junior in the title.

What to do: treat your shipping history as your main credential. If you have fewer than five years, apply to the roles that ask for fewer, and in the rest of your application lead with production work you can show.
905 roles (66%) involve building agents. 68% of those also involve evaluation work, compared with 33% of other roles. 41% of agent roles involve observability: traces, monitoring, and knowing what the agent did and why.

Evaluation is also a job title now.
There are 48 open evaluation engineer roles, with a median posted pay of $275k across the 27 that post a US range (posted ranges only, no equity or bonus).
What to do: expect an interview question about how you know your agent works, and have a real answer: a test set, a scoring method, and traces you looked at. Our guide to the feedback, verification and measurement layer around the model is a good place to start if your agents don't have this yet.
We searched 1,334 ads for the tools they name. LangChain's own 42 openings were left out so they wouldn't skew the framework count. A mention anywhere in the ad counts, "nice to have" included.
| Named in the ad | Share of 1,334 ads |
|---|---|
| Python | 53% |
| Evals | 52% |
| TypeScript | 22% |
| Kubernetes | 21% |
| LangChain | 5% |
| LangGraph | 2% |
| LlamaIndex | 1% |
| CrewAI | 1% |
82 ads (6%) name any agent framework at all. The companies naming LangChain most often are Databricks (13 ads) and Mistral (9).
Employers describe the work itself: evaluation, retrieval, tool use, production systems. Which framework you use for it rarely makes the ad.
What to do: if you have a framework on your CV, describe what you built with it, how you measured it and how it ran in production. Put the framework name second.
Across all 1,376 roles, 54% of ads name Python and 22% name TypeScript. (This base includes LangChain's ads, so the Python figure differs slightly from the table.) Among the 258 agent engineer roles, it's 48% Python and 38% TypeScript.

Agent engineers build the product around the model: the app, the tool integrations, the interface the customer sees. A lot of that is written in TypeScript, and the ads reflect it.
What to do: if you only work in Python, get comfortable enough in TypeScript to build and ship the front end of an agent. Agent engineer ads ask for it more often than any other family's.
Anthropic open-sourced the Model Context Protocol on 25 November 2024. It's a standard for connecting AI assistants to the systems where data lives. 115 ads at 45 companies now name it, and 104 of those ads are outside Anthropic.
Across the same 1,376 ads, RAG appears in 14%, reinforcement learning in 11%, fine-tuning in 10%, MCP in 8% and vector databases in 7%. A protocol less than two years old shows up about as often as vector databases.
What to do: build at least one MCP server yourself, so you can talk about tool design, auth and failure handling from experience.
57 ads at 37 companies ask for hands-on experience with Claude Code, Cursor, Codex or coding agents. We only counted companies that don't make coding tools themselves (1,047 ads), which leaves out Anthropic, OpenAI, Cursor, Cognition, Replit, Factory, Sourcegraph and Warp.
That's 5% of those ads.
What to do: the Cerebras wording describes the skill well. Use the agent, then check what it produced. Be ready to explain how you do that check.
After communication, the soft skill the ads name most is comfort with ambiguity. It appears in 582 ads at 97 companies (42%).
This counts the word anywhere in the ad, company boilerplate included.
| Named in the ad | Share |
|---|---|
| Communication | 64% |
| Ambiguity | 42% |
| Curiosity | 13% |
| Fast-paced | 11% |
| Taste | 8% |
| High agency | 8% |
Scale AI wants engineers who "translate ambiguous customer problems into scalable production AI architectures." OpenAI asks for a record of leading AI systems "from ambiguous discovery through production deployment."
What to do: prepare one story about a vague request you turned into something that shipped. Cover what you asked, what you decided, and what you cut.

Open roles by metro:
A role that lists several cities counts in each one, so these figures add up to more than the total. The Bay Area and New York together hold 54% of the 1,344 roles that state a place. London is third, with twice as many openings as Seattle. By country, the US has 64% and Europe, including the UK, has 21%.
312 roles are listed as remote. 5 of them accept applicants from anywhere. 155 remote roles are US only, 61 are limited to the UK or EU, 29 to the US and Canada, and 21 to Asia-Pacific.
We took these from each posting's own location and eligibility text.

Remote is also the smaller share. Of the 1,100 roles that state a work mode, 47% are on-site, 24% hybrid and 28% remote. Anthropic lists none of its 112 roles as remote. OpenAI lists 29 of its 141.
If you live outside the US, read the eligibility line before anything else in the ad.
OpenAI leads with 141 open roles, then Anthropic with 112. After them come Sierra (67), Wonderful (50) and Decagon (47). All three sell AI agents for customer service, and together they have 164 open AI engineering roles. Databricks (44), LangChain (42) and Scale AI (42) together have 128.
Much of that hiring is forward deployed: 26 of Sierra's 67 roles, 49 of Wonderful's 50 and 21 of Decagon's 47. These companies hire engineers to put agents in front of their customers' customers.
For a job seeker, this shows where agents are running in production today. In our data, customer service is the agent use case with the most hiring behind it. If you've built anything for support workflows, such as ticket routing, retrieval over help center content, or handoff to a human, put it near the top of your CV.
Each step ties back to a finding above.
By the end of the month you'll have one project that covers evaluation, tracing, MCP and a TypeScript interface, plus written answers to the questions these ads keep raising.
Every role in this report is listed on the agentnative.dev AI engineering jobs board. It's free to browse, and you can filter by role family, seniority and remote mode, so you can go straight to forward deployed, agent engineer or evaluation roles in the countries you can work from.