Agentic engineering

Agentic engineering is the disciplined end of the AI-coding spectrum, where AI acts as a powerful implementation engine inside carefully designed systems of constraints, tests, and feedback loops, with humans retaining oversight over architecture, correctness, and quality. Andrej Karpathy introduced the term in early 2026 to name the disciplined counterpart to vibe coding, after his original "vibe coding" framing was applied so broadly it lost meaning.

The spectrum

AI-assisted coding runs from vibe coding through structured AI-assisted coding to agentic engineering. The key differentiator is not whether you use AI but how much structure, verification, and human judgment surrounds the output.

Vibe coding

Structured AI-assisted coding

Agentic engineering

Intent specification

Casual prompts

Detailed prompts with examples

Formal specs, architecture docs, memory files

Verification

"Does it seem to work?"

Manual spot-checking

Automated test suites, CI/CD gates, LM judges

Codebase understanding

Minimal, may not read generated code

Selective review of critical paths

Comprehensive architecture review with AI on implementation details

Error handling

Copy-paste errors back to AI

Developer diagnoses root cause, AI fixes

Agents self-diagnose within bounds, humans handle architectural issues

Appropriate scope

Prototypes, scripts, hackathons

Features in established codebases

Production systems, team-scale

Risk profile

High, acceptable for disposable work

Moderate, human judgment at key checkpoints

Low, systematic verification at every stage

Tests and evals as the dividing line

The single biggest differentiator is how outputs get verified. Tests verify the deterministic parts — a function given this input produces that output. Evals verify the non-deterministic parts: did the agent take the right trajectory of steps, choose the right tools, produce a response that meets the quality bar? Evals are checked by labelled datasets, scoring rubrics, and LM judges. Without both, the practice is always vibe coding, however sophisticated the prompts. See automated testing.

The developer as factory manager

In agentic engineering the developer’s primary output is not code but the system that produces code — specifications and context, agents that translate specs into implementation, tests and quality gates, feedback loops that route failures back for correction, and guardrails that constrain agents to safe behaviour. Writing the specification before any code is generated is the substance of spec-driven development. A factory manager designs the assembly line and ensures quality control rather than assembling every widget by hand. Success comes from giving agents success criteria rather than step-by-step instructions. See agent harness and loop engineering.

The economics

Agentic engineering is high CapEx — an upfront investment in API schemas, deterministic test suites, and structured context — but low OpEx. The AI operates within a governed factory, so its output is structurally sound, pre-tested, and aligned with standards, and the marginal cost of shipping and maintaining a feature drops dramatically. This inverts vibe coding's low-CapEx/high-OpEx profile.

Context engineering is a financial lever in the token economy: a dense, high-signal payload — a precise AGENTS.md and architectural guardrails — raises first-pass success and avoids the costly trial-and-error loops of vibe coding. See context engineering and large language model.

Context engineering as the bridge

The shift from prompt engineering to context engineering is the bridge between vibe coding and agentic engineering. Models need the same context a skilled human developer would need, not cleverly worded instructions.


See also

References

  • The New SDLC With Vibe Coding, Addy Osmani, Shubham Saboo & Sokratis Kartakis (2026): From ad-hoc prompting to agentic engineering; introduces the spectrum and the factory model.