Agentic Engineering

Agentic engineering is the professional discipline of coordinating powerful, stochastic coding agents to move faster without sacrificing the quality bar expected from serious software engineering.

Key points

Karpathy distinguishes it from vibe coding: vibe coding raises the floor for everyone, while agentic engineering preserves the quality bar for professional software ^[src-055].
The engineer remains responsible for security, correctness, architecture, taste, and oversight even when agents write the implementation details ^[src-055].
The human role shifts toward detailed specs, docs, plans, durable identifiers, fundamental design choices, and review of whether the agent's output makes sense ^[src-055].
Hiring for this capability should change. Karpathy suggests evaluating large, realistic projects rather than puzzle-style coding interviews, then using other agents to attack or break the candidate's deployed system ^[src-055].
The ceiling may exceed the old "10x engineer" frame because skilled operators can coordinate many agents, tools, and verification loops at once ^[src-055].
Rory Richardson adds an operating-layer view: agentic engineering changes the whole lifecycle, compressing specs, code, operations, modernization, and review into AI Development Lifecycle workflows ^[src-057].
Richardson also stresses that AI is not deterministic. Teams should use it as an accelerant and democratizer while keeping humans responsible for verification, polish, architecture, and what ships ^[src-057].
^[src-061] adds a practitioner psychology layer: professional developers are already shipping AI-generated code, but Raschka warns that replacing all enjoyable problem-solving with agent management can erode fulfillment and agency.
The same source explains the capability jump behind coding agents: RLVR and Inference Time Scaling teach models to try tools, inspect outputs, use CLIs, navigate repos, and iterate toward verifiable success ^[src-061].
^[src-064] adds Steinberger's practitioner version: agentic engineering means empathizing with the agent's missing context, using short prompts only after architecture and files are clear, reviewing intent before implementation, and cleaning up when late-night vibe coding creates debt.
The same source shows the outer edge of agentic engineering: an agent can inspect and modify the harness that runs it, so the human's responsibility moves toward boundaries, review, security, taste, and product judgment ^[src-064].
Howell's AI-career roadmap adds a complementary hiring-skill point: many "AI engineer" jobs are closer to software engineering than ML research because they wrap and productionize existing models rather than train frontier models from scratch ^[src-075].
The AI Engineer channel corpus turns agentic engineering into an operational discipline: coding agents, specs, review, evals, context, observability, durable execution, and AI-ready codebases recur across the 2023-2026 archive ^[src-077].
The corpus also clarifies the boundary between vibe coding and professional agent use: faster code generation only becomes engineering when paired with tests, traces, quality gates, security boundaries, and product judgment ^[src-077].
Fmind's agent-skill videos reinforce the same move from chat to durable practice: an agent becomes more reliable when reusable skills and protocol knowledge are externalized into files, procedures, and integration surfaces ^[src-078].
The MLOps course adds a lower-level prerequisite for coding agents: codebases need packages, tests, configs, docs, releases, containers, monitoring, and security practices before agent acceleration is safe to compound ^[src-078].
Cursor's 2026 event adds platform telemetry to the same shift: agent requests overtook tab-completion-style interactions, 30% of Cursor's own PRs are reportedly agent-developed end-to-end, and enterprise users are increasingly delegating syntax-writing to agents ^[src-080].
The role consequence is exactly the agentic-engineering frame: humans spend more time on delegation, review, architecture, testing, and coordination across many concurrent agent workstreams ^[src-080].
OpenAI's API & Codex Build Hour adds the Harness Engineering variant: serious agentic engineering requires making the repository itself legible to agents through tests, task entry points, docs, worktrees, skills, standards, and persistent decision notes ^[src-084].
The same source says decision-making can become the bottleneck once agents write code quickly, so teams need specs, notes, synchronous human sync, and review rituals that keep architecture and product intent coherent ^[src-084].
The late-May AI Engineer update adds the evidence-gated version of this discipline: Nick Nisi's WorkOS talk says agents should be required to prove tests, UI fixes, and completion evidence through harness gates rather than promises in text ^[src-088].
The same update adds a skills lesson: compact, measured gotchas can outperform comprehensive generated skill dumps when they preserve focus and avoid context noise ^[src-088].
^[src-094] formalises the spectrum from Vibe Coding to agentic engineering: the differentiator is the structure, verification, and human judgement around the AI's output.
The paper makes tests and evals the boundary between serious agentic engineering and advanced-looking vibe coding: tests verify deterministic behaviour, while evals verify trajectories, tool choices, and final-response quality ^[src-094].
It also adds the economic frame: agentic engineering has higher upfront investment, but can reduce token burn, maintenance tax, and security remediation through reusable context, tests, harnesses, and routing ^[src-094].
^[src-191] adds the World's Fair Online Track version: serious agentic engineering now includes review-debt management, bug triage, save/resume state, liveness semantics, multi-agent factories, skills as reusable capability packages, and CI/CD-like controls around agent work.
The update also makes receipts a quality bar: coding agents should leave inspectable evidence, not only diffs or claims of completion ^[src-191].

Related entities

Related concepts

Source references

^[src-055] Sequoia Capital — "Andrej Karpathy: From Vibe Coding to Agentic Engineering" (2026-04-29)
^[src-057] Amazon Web Services — "The Future of Agentic AI with Rory Richardson | AWS Humans In The Loop Podcast" (2026-05-01)
^[src-061] Lex Fridman – "State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490" (2026-01-31)
^[src-064] Lex Fridman – "OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger | Lex Fridman Podcast #491" (2026-02-12)
^[src-075] Egor Howell — "STOP Taking Random AI Courses – Read These Books Instead" (2025-06-14)
^[src-077] AI Engineer channel transcript cluster (678 saved transcripts, 2023-10-20 to 2026-05-15)
^[src-078] Mederic Hurier (Fmind) channel transcript cluster (62 saved transcripts, 2024-11-26 to 2026-05-14)
^[src-080] Cursor — "The next era of AI coding" (2026-05-12)
^[src-084] OpenAI Codex, Workspace Agents, Prompt Caching, and Superintelligence Policy cluster (2026-02-09 to 2026-05-08)
^[src-088] AI Engineer late-May 2026 channel update (48 transcripts, 2026-05-15 to 2026-05-31)
^[src-094] Addy Osmani, Shubham Saboo, Sokratis Kartakis – "The New SDLC With Vibe Coding" (2026-05)
^[src-191] AI Engineer World's Fair Online Track 2026 playlist update (47 new transcript captures, 2026-06-22 to 2026-07-02)

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 491 indexed pages and articles.