Skip to content
Code Guide

AI Roles & Career Paths: The New Engineering Landscape

AI Roles & Career Paths: The New Engineering Landscape

Section titled “AI Roles & Career Paths: The New Engineering Landscape”

Last updated: March 2026

Confidence: Tier 2 — Based on job market data, industry publications, and emerging field research

Reading time: ~20 minutes

The AI wave didn’t just create new tools. It created new jobs that didn’t exist 3 years ago and is reshaping what existing roles mean. This guide maps the full landscape: what each role does, what skills it requires, how they relate to each other, and where each one is heading.


  1. The Landscape in One View
  2. Prompt Engineer
  3. Context Engineer
  4. AI Engineer
  5. LLM Engineer
  6. AI Agent Engineer
  7. Founding AI Engineer
  8. AI Architect
  9. Platform Engineer (AI context)
  10. Harness Engineer
  11. AI Product Manager
  12. AI Safety & Eval Engineer
  13. ML Engineer
  14. Career Decision Matrix
  15. Salary Benchmarks (2025-2026)
  16. What’s Not a Role (Yet)
  17. Job Listings

Two axes structure this landscape: proximity to the model (are you training it, prompting it, or building infrastructure around it?) and proximity to production (research vs. shipped product).

← Closer to the model Closer to infrastructure →
Research ML Engineer ←────────────────────────── AI Architect
AI Safety Engineer Platform Engineer
│ │
│ │
Production LLM Engineer ──── AI Engineer ──── AI Agent Engineer
Context Engineer Harness Engineer
Prompt Engineer Founding AI Engineer
AI Product Manager

Most new job demand sits in the bottom-right: building reliable AI systems that ship and stay reliable in production. The “pure research” quadrant remains competitive and specialized. The highest growth is in the applied, product-facing roles.


Status: First wave (2022-2023), partially commoditized but still relevant in specialized contexts.

Craft and optimize the instructions sent to AI models to get reliable, high-quality outputs. The scope ranges from one-shot prompts to complex multi-step prompt chains for production systems.

  • Design prompt templates for specific use cases (customer support, code generation, document analysis)
  • Run systematic A/B tests to measure prompt performance
  • Document prompt libraries and version them
  • Optimize prompts for cost (fewer tokens, same quality)
  • Work with domain experts to encode knowledge into prompts
TechnicalSoft
Understanding of LLM behavior and failure modesCommunication with non-technical stakeholders
Basic Python (for automation and testing)Systematic experimentation mindset
Familiarity with evaluation frameworksAttention to edge cases
Versioning practicesDocumentation discipline

The “prompt engineer” title as a standalone role is consolidating into broader AI Engineer or Context Engineer roles. Where it persists: companies with very specific, high-stakes prompt domains (legal, medical, financial compliance). Upskill toward context engineering or AI engineering if you’re in this role.

Technical writer, QA engineer, domain expert (law, medicine, finance), content strategist.


Status: Emerging — one of the fastest-growing specializations in 2025.

Context engineering is the evolution of prompt engineering. Where prompt engineers craft instructions, context engineers design systems that give AI models the right information, at the right time, in the right format. Andrej Karpathy explicitly moved from “vibe coding” framing to “context engineering” as the more precise description of this work.

“Context Engineering is providing the right information and tools, in the right format, at the right time.” — Philipp Schmid, Google

  • Design RAG (Retrieval-Augmented Generation) systems and knowledge bases
  • Manage context windows across multi-turn interactions and long-horizon tasks
  • Define what agents remember, retrieve, or forget during task execution
  • Structure information hierarchies (system prompts, conversation history, retrieved docs, tool definitions, safety constraints)
  • Optimize context for accuracy and cost simultaneously
  • Measure context quality through systematic evals
TechnicalSoft
Python (context pipeline automation)Systems thinking
Vector databases (Pinecone, Chroma, Weaviate)Information architecture instinct
SQL and NoSQL (context retrieval)Cross-functional collaboration
Cloud platforms (AWS/Azure/GCP)Curiosity and continuous learning
RAG architectures, embedding modelsPrecision in documentation

Context engineers work upstream of AI engineers (they define what context is available) and downstream of domain experts (they encode domain knowledge into retrievable structures). Closely related to platform engineers in large organizations.

Data engineer, backend engineer, ML engineer, information architect.


Status: Mainstream — the generalist role for building AI-powered products.

Build end-to-end AI systems. Not researchers (they don’t train models from scratch), but not just integrators either. They take LLMs and orchestration frameworks and build systems that ship. Think of them as software engineers who’ve added LLM integration, evals, and AI product intuition to their stack.

  • Design and implement LLM-powered applications (chatbots, agents, pipelines)
  • Build evaluation frameworks to measure model output quality
  • Integrate AI capabilities into existing software systems
  • Monitor AI systems in production (latency, cost, quality drift)
  • Select appropriate models for specific tasks (capability vs. cost tradeoffs)
  • Implement fine-tuning or RAG when base models aren’t sufficient
TechnicalSoft
Strong software engineering foundationsProduct judgment
Python (primary), JavaScript (often needed)Pragmatism over research purity
Familiarity with major LLM APIs (Anthropic, OpenAI, Gemini)Fast iteration mindset
Eval design and measurementAbility to work with ambiguous requirements
Understanding of embeddings, RAG, agent frameworksCommunication of AI limitations to stakeholders
MLOps basics (deployment, monitoring, versioning)

AI engineers work with existing models. ML engineers build and train models. In practice, most companies hiring in 2025-2026 need AI engineers (apply the models) not ML engineers (build the models).

Software engineer (most common), backend engineer, data engineer, ML engineer transitioning to applied work.


Status: Specialized variant of AI Engineer, prominent in model-heavy companies.

Deep specialization in large language model integration and optimization. Where AI engineers are generalists, LLM engineers go deep on the model layer: fine-tuning, RLHF, model selection, prompt optimization at scale, and evaluation infrastructure.

  • Fine-tuning base models for domain-specific tasks
  • Designing and running systematic model evaluations (evals)
  • Implementing RLHF or similar feedback mechanisms
  • Model performance benchmarking and regression testing
  • Managing model versions and A/B testing new model releases
  • Building tooling for model monitoring and drift detection
TechnicalSoft
Python (fluent)Scientific rigor
PyTorch or JAXStatistical thinking
Transformers architecture knowledgePatience with slow feedback loops
Evaluation framework designDocumentation of experiments
Distributed training basics

Strong demand at AI companies (Anthropic, OpenAI, scale-ups) and in large enterprises building proprietary models. Distinct from AI engineer in its proximity to the model itself. Expect this role to bifurcate: pure research at labs vs. applied fine-tuning at enterprises.


Status: High growth — one of the most in-demand specialized roles in 2025-2026.

Design and build autonomous agent systems. While AI engineers build general AI products, agent engineers specialize in systems that plan, reason, use tools, and execute multi-step tasks without constant human intervention.

  • Design multi-agent architectures (orchestrator + specialist agents)
  • Build agent memory systems (short-term, long-term, episodic)
  • Implement tool use and API integrations for agents
  • Design guardrails and safety mechanisms for autonomous systems
  • Build human-in-the-loop checkpoints for high-risk decisions
  • Monitor agent behavior in production (reliability, cost, anomaly detection)
  • Test agent systems systematically (agentic eval is a distinct discipline)
TechnicalSoft
Agent frameworks (LangChain, AutoGen, Claude Agent SDK, CrewAI)Systems thinking
Orchestration patternsRisk judgment (when to let agents act autonomously)
Tool/API integrationUser experience intuition
Async programmingDebugging patience (agents fail in non-deterministic ways)
Observability and tracing (LangSmith, Langfuse, etc.)

Non-determinism. Agent systems fail in ways that are hard to reproduce. Observability tooling (tracing every agent step) is as critical as the agent code itself. Engineers who treat agent debugging like debugging traditional code struggle.


Status: Highly sought after in AI-native startups and seed-to-Series A companies.

A hybrid role unique to early-stage companies: part AI engineer, part product engineer, part technical co-founder. They own core product functionality end-to-end, from architecture decisions to customer interactions, while building on top of AI capabilities.

Typically targets engineers with 0-4 years of experience who are comfortable with ambiguity, figure things out independently, and already use AI tools daily in their workflow.

  • Build entire product features from architecture to deployment, not just assigned tickets
  • Make foundational technical decisions that will shape the company’s stack for years
  • Work directly with founders on product strategy and prioritization
  • Use AI coding tools as force multipliers to ship at startup speed
  • Interact directly with early customers to understand problems
  • Define engineering culture before it calcifies

Scope of ownership and ambiguity. A senior engineer at a large company works within defined systems. A founding engineer defines the systems. The leverage is massive in both directions: great decisions compound, bad ones become technical debt that’s hard to escape.

  • Bias toward action over analysis paralysis
  • Comfort shipping imperfect things and iterating
  • Product intuition alongside technical skills
  • Already fluent with AI coding tools (Claude Code, Cursor, Copilot)
  • Able to context-switch from infra to product to customer research in the same day

Strong mid-level engineers at established companies who want more ownership. Common source: engineers who’ve been quietly building side projects with AI tools.


Status: Senior/Staff level — emerging role in larger organizations.

Design enterprise AI systems at the system level. Where AI engineers ship features, AI architects define the patterns, platforms, and decision frameworks that multiple teams use. They make the technology choices that others live with for years.

  • Define AI technology strategy and stack decisions (which models, which frameworks, which providers)
  • Design enterprise AI reference architectures
  • Set standards for AI system observability, security, and governance
  • Evaluate build vs. buy decisions for AI capabilities
  • Ensure AI systems are scalable, cost-effective, and auditable
  • Bridge between business requirements and technical AI implementation
  • Deep experience across AI/ML stack (models, infrastructure, MLOps)
  • Strong communication skills (presenting to C-suite, working with legal/compliance)
  • Understanding of cloud provider AI offerings (AWS Bedrock, Azure OpenAI, Vertex AI)
  • Security and compliance awareness (GDPR, AI Act, SOC2)
  • Experience designing distributed systems at scale

Senior AI engineer → Staff → Architect. Often takes 5-8 years in AI-adjacent roles. Alternatively: cloud architect + strong AI self-study.


Status: Established role, significantly reshaped by AI.

Build and maintain the internal developer platform. With AI, this role has expanded to include the “golden path” for AI development: standardized ways for teams to integrate LLMs, common observability infrastructure, cost controls, and guardrails so individual teams don’t reinvent the wheel or create security risks.

AI-specific responsibilities added to traditional platform work

Section titled “AI-specific responsibilities added to traditional platform work”
  • Provide standardized LLM integration patterns (internal SDKs, proxies, abstractions)
  • Manage API keys, rate limits, and cost allocation across teams
  • Build AI observability infrastructure (tracing, logging, alerting)
  • Enforce security policies for AI outputs (PII filtering, output validation)
  • Maintain model registries and versioning systems
  • Create “paved roads” for RAG patterns, agent architectures, eval pipelines

When every team is building their own LLM integrations, you get: duplicated cost, inconsistent security, no centralized observability, and no shared learnings. Platform engineers who understand AI prevent this fragmentation. They’re the reason the AI investment in a company scales instead of sprawling.

MLOps tooling, LLM gateway products (LiteLLM, Portkey), cloud AI services, cost optimization patterns, security for AI (prompt injection mitigation, output filtering).


Status: Emerging — formalized by Martin Fowler in 2025, not yet institutionalized as a standalone title.

Build the infrastructure that keeps AI agents “under harness” — under control. As agentic AI systems generate code, take actions, and operate with increasing autonomy, harness engineers build the systems that ensure they stay within architectural constraints, produce coherent output, and don’t accumulate entropy over time.

Source: Martin Fowler — Harness Engineering

1. Context engineering (knowledge infrastructure) Not one-off prompts, but a continuously updated knowledge base embedded in the codebase. Agents know your conventions, architecture decisions, and domain context. Dynamic access to observability data and documentation.

2. Architectural constraints (agent guardrails)

  • LLM-based watchdog agents that review generated code
  • Custom deterministic linters enforcing your specific architectural patterns
  • Structural tests (ArchUnit-style) that run automatically
  • Pre-commit hooks that reject code violating established constraints

3. Entropy management (drift prevention) Periodic agents that scan the codebase for: outdated documentation, architectural violations that slipped through, abandoned patterns that reappeared, inconsistencies introduced by multiple agents working in parallel.

Without a harness, AI agents produce code that individually looks fine but collectively drifts away from your architecture, your patterns, and your documentation. The harness is what makes “AI generates most of the code” sustainable at scale rather than a path to unmaintainable systems.

This role pushes toward intentional technological convergence: organizations with 2-3 primary tech stacks benefit far more from standardized harnesses than organizations with 10 different stacks. It’s a deliberate trade of technical freedom for reliability.

“Ce n’est pas quelque chose dans lequel vous pouvez vous lancer pour des résultats rapides.” — Martin Fowler

Currently absorbed by: platform engineers, staff/principal engineers, architecture guilds. Likely to become an explicit role in:

  • Companies running autonomous coding agents at scale
  • Large enterprises with 50+ engineers using AI coding tools
  • Organizations that’ve experienced “AI entropy” firsthand (code that works but nobody understands anymore)

Software architecture, linter/static analysis tooling, LLM orchestration, observability, codebase knowledge management, entropy detection patterns.


Status: Mainstream and growing, with significant premium over traditional PM roles.

Product management with deep AI fluency. They understand what AI can and can’t do, manage the unique product challenges of AI systems (non-determinism, latency, hallucinations, cost), and translate between business needs and AI capabilities.

  • Define product requirements for AI features with technical constraints in mind
  • Work with AI engineers on evaluation criteria (what does “good” look like?)
  • Manage the unique UX challenges of AI: uncertainty, latency, error handling
  • Own the cost/quality/speed tradeoffs for AI features
  • Communicate AI limitations and risks to stakeholders
  • Run A/B tests on model versions, prompt changes, feature changes

What makes AI PM different from traditional PM

Section titled “What makes AI PM different from traditional PM”

Traditional PM ships features that behave deterministically. AI PMs ship systems where outputs vary. They need to think probabilistically: not “will this work?” but “what % of the time will this work, and what happens in the other cases?” Quality measurement is continuous, not binary.

Standard PM skills (roadmapping, prioritization, user research) plus: LLM API familiarity, eval design, basic Python for running experiments, understanding of model tradeoffs (accuracy vs. cost vs. latency), AI UX patterns.

FAANG-level: $160K-$200K+ entry-level AI PM. Senior: $200K-$300K+ total compensation.


Status: Specialized — primarily at AI labs and companies with regulated AI deployments.

Ensure AI systems behave safely, reliably, and in alignment with intended values. Two related but distinct specializations: Eval Engineers (build systems to measure model behavior) and AI Safety Engineers (identify and mitigate risks in AI systems).

  • Design evaluation frameworks (evals) to measure model quality, safety, and capabilities
  • Build automated eval pipelines that run on every model version change
  • Define metrics that capture real-world performance (not just benchmark gaming)
  • Implement human evaluation workflows for subjective quality dimensions
  • Detect regressions before they reach production
  • Red-team AI systems to find failure modes, jailbreaks, and harmful outputs
  • Implement content filtering, output validation, and guardrail systems
  • Design human-in-the-loop checkpoints for high-risk decisions
  • Monitor production systems for harmful outputs or unexpected behavior
  • Work with legal/compliance on AI governance

Rigorous experimental design, statistics, Python, strong understanding of LLM failure modes, communication skills for risk reporting.

Primarily: Anthropic, OpenAI, Google DeepMind, Meta AI, Microsoft AI. Growing in: healthcare, finance, legal tech — regulated industries where AI errors have serious consequences.


Status: Established — the most traditional of the AI engineering roles.

Develop, train, deploy, and maintain machine learning models. In the LLM era, many ML engineers have pivoted toward fine-tuning and applied AI work rather than building models from scratch — that work is increasingly concentrated at a small number of frontier labs.

  • Data pipeline development (collection, cleaning, transformation)
  • Model training and fine-tuning
  • Feature engineering
  • Model serving and deployment (MLOps)
  • Performance optimization and model compression
  • Production monitoring for model drift

The “build a model from scratch” path is increasingly rare outside frontier labs. ML engineers in most companies now work on: fine-tuning existing models, building RAG systems, deploying and monitoring models in production, and bridging between AI engineers and data infrastructure. The practical overlap with AI engineer is large.

Python (fluent), PyTorch or TensorFlow, distributed computing, data pipeline tools (Spark, Airflow, dbt), cloud ML platforms (SageMaker, Vertex AI, Azure ML), statistical foundations.


Which role fits your current background and goals?

Your current profileBest next roleTimeline
Software engineer (3+ years) who wants to work with AIAI Engineer3-6 months upskill
Software engineer at early startup who wants ownershipFounding AI EngineerNow, if opportunity exists
Backend engineer interested in infra + AIPlatform Engineer (AI)6-12 months
Senior engineer who thinks in systemsAI Architect or Harness Engineer1-2 years experience accumulation
Engineer who likes research and rigorLLM Engineer or AI Safety/Eval+ML foundations needed
Non-technical who works with AI dailyPrompt Engineer → Context Engineer6-18 months
PM who wants to stay PM but be more relevantAI Product Manager3-6 months upskill
Engineer obsessed with reliability and architectureHarness Engineer (emerging)Pioneers’ territory

The fastest path to AI employment in 2025-2026

Section titled “The fastest path to AI employment in 2025-2026”
  1. Build something with AI APIs (Claude, OpenAI) — a real project, not a tutorial
  2. Write about what you built (blog post, GitHub README, LinkedIn)
  3. Add evaluation: measure your system’s quality, show the numbers
  4. Apply for AI Engineer roles — the bar is demonstrated building, not credentials

Note: 76% of candidates claiming AI expertise lack production-level deployment experience (LangChain State of Agent Engineering 2025). The bar is lower than it appears if you’ve actually shipped something.


Indicative only — large variance applies. These figures are US market base salaries (2025-2026). Europe runs 30-50% lower, other markets 40-60% lower. Total compensation (equity, bonus, RSUs) can significantly exceed base, especially at startups and FAANG. Experience level, location within a country, company stage, and negotiation all create wide variance. Use these as orientation, not negotiation anchors.

RoleEntryMidSeniorNotes
Prompt Engineer$80K-$110K$110K-$150K$150K-$180KShrinking standalone market
Context Engineer$100K-$140K$140K-$180K$180K-$230KGrowing fast
AI Engineer$120K-$160K$160K-$220K$220K-$300KHighest volume of open roles
LLM Engineer$130K-$170K$170K-$250K$250K-$350KLab-level roles higher
AI Agent Engineer$130K-$170K$170K-$240K$240K-$320KStrong demand 2025-2026
Founding AI Engineer$100K-$150K + equityEquity makes total comp wide-ranging
AI Architect$180K-$260K$260K-$380KSenior/Staff only
Platform Engineer (AI)$110K-$150K$150K-$210K$210K-$280K
Harness EngineerNot yet standardizedAbsorbed into other roles
AI Product Manager$130K-$170K$170K-$230K$230K-$350KFAANG premium significant
AI Safety/Eval Engineer$140K-$180K$180K-$250K$250K-$400KLab compensation highest
ML Engineer$100K-$140K$140K-$200K$200K-$280KLower demand outside labs

Sources: FinalRoundAI (2025), Alcor AI Salary Report (2025), RiseWorks AI Talent Report (2025), job postings analysis.


Some terms you’ll hear that describe practices or methodologies, not job titles:

Vibe coder — A methodology (use AI coding assistants to handle implementation while you focus on design), not a job. Andrej Karpathy coined the term then himself pivoted toward “context engineering” as more precise. No serious company has “Vibe Coder” on a job description.

AI-native engineer — Describes a quality expected of all engineers increasingly, not a specialized role. It means: you use AI tools fluently in your daily workflow. It’s the bar, not the title.

Orchestration engineer — Sometimes used for agent systems, overlaps significantly with AI Agent Engineer. Not yet a distinct category.

These terms are worth knowing (you’ll encounter them in job descriptions and articles) but don’t represent distinct career paths — yet.


Coming soon — Curated listings for AI roles at companies building seriously with Claude Code and agentic AI.

If you’re hiring for any of the roles described in this guide, reach out to discuss featuring your opportunity here.