Back to Blog

From Full-Stack to AI Engineer: How the Elite Code in 2026 and My 3-Year Plan

Angel Valladares14 min readPublished on September 5, 2026
AI Engineering
Career
Claude Code
Agents
Evals

I have been a full-stack developer for years, and I have never seen the industry move this fast. The engineers I admire no longer spend their day typing code: they run several AI agents in parallel and spend most of their time planning and reviewing. At the same time, a new role, the "AI engineer", is absorbing much of what used to be product development.

So I spent a day doing proper research: roughly eighty sources, prioritizing June to September 2026, and skeptical of anything that smelled like vendor marketing. This post is the result. It has three parts: how the elite actually work today, what an AI engineer is and gets hired for, and the plan I am following to get there over the next one to three years. If you are a web developer wondering where to point yourself, I hope it saves you some of the uncertainty I felt.

The verdict in five points

  1. The elite barely write code anymore; they plan, delegate, and verify. The dominant workflow is plan → 3–5 agents in isolated git worktrees → automatic verification (tests, CI, browser) → risk-based review → merge. But only 31% of professional developers are "agentic coders" (JetBrains, Aug 2026), so mastering this flow today is still a real advantage.
  2. "AI engineer" in 2026 means an engineer who builds products on top of models via API. RAG, agents with tools, evals, MCP servers, observability, and process automation. It is not ML research and it does not require training models.
  3. Demand is real and measurable. AI Engineer is the fastest-growing role in the US according to LinkedIn (Jan 2026); there were 43,480 US "AI engineering" postings between January and July 2026 with a median posted salary of $176K; Forward Deployed Engineer postings grew 729% year over year (Indeed, Apr 2026).
  4. The gap for a full-stack developer is narrower than it looks, but concrete. We already have what 95% of postings ask for as a base. What is missing: evals and error analysis (the #1 interview signal), retrieval quality, context engineering, bounded agent design, cost and latency economics, LLM security, and "production-grade" Python.
  5. The best-fitting path for a JS/TS developer is "AI product engineer". It is the profile Stripe, Vercel, and product startups post (TypeScript accepted, Python as a second language), it is more often remote, and it is the natural door into Applied AI.

How the elite code in September 2026

The tools, by weekly use

JetBrains surveyed more than 15,000 professional developers between May and July 2026: 90% use AI coding agents at least weekly, and 68% daily. Claude Code leads with 39% weekly use, followed by GitHub Copilot at 21% (down from 29%), OpenAI Codex at 16% (up from 3% in January), Cursor at 12% (down from 18%), OpenCode at 7%, and Antigravity at 6%. Among senior engineers the skew is stronger: in the Pragmatic Engineer survey (March 2026, median 11–15 years of experience) Claude Code is both the most used and the most "loved" tool, and 63.5% of Staff+ engineers use agents regularly.

The workflow that repeats in every first-hand account

Boris Cherny (creator of Claude Code), OpenAI's "harness engineering" team, Mitchell Hashimoto, Simon Willison, Thorsten Ball, and Addy Osmani describe, with variations, the same working day:

  1. Queue background work first thing. Three to five tasks in cloud agents (Codex cloud, Claude Code web, Cursor cloud): maintenance, research, library comparisons, migrations.
  2. Specify before implementing. Plan mode or a spec.md iterated with the agent until the plan is convincing; often a fresh session critiques the plan "as a senior engineer" before approval (the Research → Plan → Implement → Review pattern).
  3. Launch 2–5 local agents in isolated worktrees. One git worktree per agent and per orthogonal task, named sessions, notifications when one needs input. Cherny runs about five terminals plus 5–10 web sessions; Osmani warns that 3–4 in parallel is already "mentally taxing".
  4. Every agent gets a verification loop. Anthropic's number one tip: tests (red/green TDD), curl against the API, a browser through Chrome DevTools MCP, lint and typecheck hooks. "Tests are no longer even remotely optional" (Willison).
  5. AI review before human review, and human review by risk. A different model in a fresh session reviews first; the human reads line by line only for auth, payments, secrets, or "when the blast radius would be huge" (Ball). PRs of 250 lines or fewer are a common gate.
  6. Encode taste into the repo. A short CLAUDE.md / AGENTS.md (about 100 lines, an index into docs/), skills with exit criteria, custom linters that enforce the architecture, and updating those files with every mistake the agent makes.
"If I'm coding, I want an agent planning. If they're coding, I want to be reviewing." — Mitchell Hashimoto, February 2026
"Humans steer, agents execute. Corrections are cheap, and waiting is expensive." — OpenAI, "Harness engineering", February 2026. A team of 3→7 engineers shipped ~1M lines and ~1,500 PRs in five months, all written by Codex.

What it is costing (the part nobody puts in the pitch)

Faros AI's telemetry across 22,000 developers (April 2026) shows median time in review up 441%, incidents per PR up 243%, and bugs per developer up 54%. METR's follow-up study (February 2026) could not distinguish the productivity gain from zero and declared its own design compromised. DORA (May 2026) measures 35–40% gains on simple tasks but 10% or less on complex legacy code, and calls the dip the "J-curve" and the "instability tax". On security, 2026 brought a prompt-injection CVE in Claude Code's GitHub Action, the "Clinejection" supply-chain compromise, and "slopsquatting": about 20% of the packages agents recommend do not exist.

The takeaway for me: the scarce skill is not "using Claude Code". It is designing the harness (verification, review, permissions, docs, evals) that makes five agents produce code you can defend in production. That is what Anthropic, OpenAI, and the AI Engineer World's Fair now call harness engineering, and it is the first phase of my plan.

What an AI engineer actually is

swyx coined the term in 2023 as "the engineer on the application side of the API line": to the left of the line, people train models; to the right, people build products with them. By 2026, the AI Engineer World's Fair (6,000+ attendees) and Andrew Ng's "AI Engineering Skills Map" (August 2026) treat it less as a niche and more as where all software engineering is heading. Ng's four skill areas: building and deploying AI applications, software engineering fundamentals, using coding agents, and "shaping the build" (product judgment).

The titles, decoded

  • AI Engineer / LLM Engineer: builds apps on foundation models (RAG, agents, evals, serving). Python in 62% of postings.
  • AI Product Engineer / Full-stack AI Engineer: a full-stack developer whose features are LLM-powered; owns UX, backend, prompts, and evals. Next.js/TS plus Node or Python. Posted by product startups, the Vercel ecosystem, Stripe.
  • Applied AI Engineer: same as AI engineer but customer-adjacent (labs) or embedded in product.
  • Forward Deployed Engineer (FDE): embeds with a customer and ships LLM apps in their environment; about 60% customer-facing, 25–50% travel. Palantir, OpenAI, Anthropic, Databricks, Sierra, Harvey.
  • Agent Engineer: builds the agents themselves: tool use, planning loops, memory, evals.
  • AgentOps / LLMOps / AI Platform: runs agents in production: tracing, eval pipelines, cost, latency, guardrails.
  • ML Engineer: trains, fine-tunes, and serves models. A different job.

What they build and what interviews test

From real postings: RAG and search over company documents, internal copilots and back-office automation over CRM/ERP systems, document extraction pipelines (where much of the enterprise value is, not chat), customer support and voice agents, coding-agent harnesses, and vertical copilots for legal, health, and finance. Stripe's current "AI Engineer" posting is a good template: "ship code daily", build internal tools and agent-driven automation, know two of Ruby/Node/Python/Next.js, and use AI tools in your own workflow. A full-stack JS/TS developer already qualifies.

Interviews in 2026 weigh four things: evals (moving "beyond vibes" to golden datasets, LLM-as-judge validated against human labels, regression suites gated in CI), naming RAG failure modes unprompted, judgment about when not to build an agent, and cost and latency reasoning ("halve the cost of this feature without degrading quality"). And the question that opens almost every senior interview: "what has gone wrong for you?"

The market, and remote from Honduras

Of those 43,480 US postings, 48% are hybrid, 32% fully remote, and 20% on-site; 66% are individual-contributor roles. The largest hiring sector is professional services (28%), ahead of technology companies (24%): 76% of enterprise AI use cases are bought rather than built (Menlo Ventures), so somebody has to integrate them into real processes. Pragmatic Engineer (May 2026) reports big tech has 50–100% more AI engineering listings than a year ago, and that AI engineering compensation now exceeds software engineering compensation.

The uncomfortable but useful part: Honduras does not appear in any LATAM AI-talent guide (Mexico, Argentina, Brazil, Colombia, and Costa Rica do). A senior "LATAM remote" AI engineer is quoted at roughly $66K–96K, versus $145K–190K for US startup roles that accept contractors abroad, and $200K–300K+ for lab FDE roles that mostly require US/UK/EU presence. The objective advantage from here is the time zone: UTC-6 is US Central, which every nearshore guide lists first. The overall software market is soft but "tilting toward seniority" (Indeed Hiring Lab, July 2026).

"Run towards the waves, and get AI on your resume, immediately… you'll be filtered out if you don't have AI experience." — Charity Majors, August 2026

My gap, in order of importance

What I already have and postings require: production shipping, APIs, cloud and CI/CD, TypeScript/Node/React/Next.js, internal tools and workflow automation, fluency with coding agents, and product judgment. What I am missing, ordered by how often hiring sources cite it:

  1. Evals and error analysis.
  2. Retrieval quality, not just "RAG": hybrid search, reranking, failure modes.
  3. Context engineering: intentional compaction, the long-context "dumb zone".
  4. Bounded agent design, and knowing when a workflow beats an agent.
  5. Cost and latency economics: model routing, caching, structured outputs.
  6. Security: prompt injection, exfiltration through tools, the OWASP LLM Top 10.
  7. Observability: tracing, prompt versioning, drift, cost monitoring.
  8. Production-grade Python: FastAPI, Pydantic, async.
  9. Document pipelines: chunking, OCR, extraction, deduplication.
  10. Fine-tuning basics (low priority): knowing when to choose it over RAG or prompting.

The roadmap, 2026–2029

Phase 0 · September–October 2026 · Work like the elite at my current job

Zero cost, immediate return, and the first line of an AI engineer résumé. Adopt plan → worktrees → verification → risk-based review on a real project; write and maintain the repo's CLAUDE.md; set up hooks, permission allow-lists, and a /commit-push-pr command; add AI review before human review and measure time-in-review for four weeks. Training: Anthropic Academy (Claude Code in Action, MCP intro and advanced, Agent Skills and Subagents) and Anthropic's engineering posts on effective agents, context engineering, evals, and harness design. All free.

Phase 1 · November 2026–February 2027 · Fundamentals in my stack + first real project

Anthropic Academy's "Building with the Claude API" (8 hours), Matt Pocock's AI SDK v6 Crash Course (the TypeScript AI course: agents, tool calling, eval-driven development, guardrails), Chip Huyen's AI Engineering, production Python with FastAPI and Pydantic, and the MCP 2026-07-28 spec. Project 1: RAG over the documents of a business I help, with hybrid retrieval, reranking, citations, and a 50-question eval set, instrumented with Langfuse from day one.

Phase 2 · March–June 2027 · Evals and agents, the differentiator

Hamel Husain's free evals material and his evals-skills repo, the W&B and DeepLearning.AI eval courses, Andrew Ng's Agentic AI course, OpenAI's "Building Agents" track for breadth, and the Claude Certified Developer exam. Projects 2–4: an eval harness with a labeled failure taxonomy and a CI regression gate; an MCP server for a real system with OAuth 2.1; and a back-office agent built with the Claude Agent SDK, with approval hooks and documented cost per run. AI Engineer Miami (April 26–27, 2027) is the closest industry event.

Phase 3 · July–December 2027 · Depth, visibility, first applications

Hugging Face's Agents and Context courses, LangGraph's intro for durable-state concepts, Google's agents intensive. Projects 5–7: a document extraction pipeline with a human review UI, a voice agent for a hackathon, and an internal copilot with cost and latency dashboards. Contribute to MCP, Langfuse, Promptfoo, or Mastra; publish three technical posts with real numbers; start applying to remote AI product engineer roles.

Years 2–3 · 2028–2029 · Specialty and senior profile

Stanford CS336 and Karpathy's nanochat to understand models from the inside, Raschka's two "from scratch" books, the Berkeley Agentic AI MOOC, the Claude Certified Architect track if I move toward architecture or consulting, and a conference talk about a system of mine in production.

Budget

The core costs about $675–885 over twelve months: the AI SDK course ($149), Huyen's book (~$50), two months of DeepLearning.AI Pro (~$50–60), the CCDV-F exam ($125), and $300–500 in API credits for seven projects with eval suites. The $4,200 evals cohort on Maven is out of budget and its free material covers most of it; cloud certifications only if an employer's stack demands them.

The seven portfolio projects

In this order, each with an eval suite and traces, because that combination is the portfolio:

  1. RAG over a real business's documents (Next.js, Vercel AI SDK, pgvector, Promptfoo).
  2. An eval harness for that RAG, with an LLM judge validated against human labels.
  3. An MCP server for a real system, OAuth 2.1, published to a registry.
  4. A back-office agent: invoices → extraction → validation → ERP, with approval hooks.
  5. A document extraction pipeline with confidence scores and a human review UI.
  6. A voice agent with tool calling.
  7. An internal copilot with OpenTelemetry GenAI spans and cost/latency SLOs.

What not to believe

  • "Stack Overflow 2026 results": the survey opened June 23 and has not published results; articles quoting "2026" numbers recycle 2025 data.
  • "A million lines in a week" and fleets of 30 agents: the teams with verifiable throughput credit linters, docs, tests, and sandboxes, not autonomy.
  • Self-reported adoption figures: "98% of OpenAI employees use Codex" and "47% of code is agent-written" are perception or vendor data; independent telemetry shows the cost side.
  • $450K+ FDE salaries: frontier labs only, with 55–70% in equity. FDE and AI roles at ServiceNow, Stripe, or startups sit at $130K–235K base.
  • "Software development will stop mattering": what depreciates is code as the bottleneck. What appreciates is system design, verification, product judgment, and integration into business processes. swyx: "LLMs reward expertise; they raise the ceiling much more than they raise the floor."

I will update this post as I move through the phases. If you are on a similar path, I would love to compare notes.

Sources

From Full-Stack to AI Engineer: How the Elite Code in 2026 and My 3-Year Plan — Angel Valladares