Agentic Engineering: Disciplined Work With Coding Agents

Agentic AIArchitecture and orchestrationPublished By Simon Budziak

Agentic engineering is the practice of building production software with AI coding agents under real engineering discipline. Where vibe coding accepts generated code on vibes, agentic engineering specifies the work, lets agents implement against tests, and reviews before anything ships.

How is agentic engineering different from vibe coding?

One word: verification. Vibe coding is conversational and trusting: describe, accept, run, repeat. It is excellent for prototypes and disposable tools and dangerous for anything that handles money or data. Agentic engineering keeps the agent’s speed and restores the controls: a written spec, tests the agent must pass, and human review before merge. Simon Willison, who popularized the framing, defines it simply as developing software with the assistance of coding agents, tools like Claude Code, Codex and Gemini CLI (Simon Willison, Agentic Engineering Patterns, read 12 October 2026).

What does the agentic engineering workflow look like?

Spec first: the engineer writes what the change must do and how success is checked. Then delegation: a coding agent implements, runs the tests, fixes its own failures, often with subagents fanned out for research or review. Then the human pass: reading the diff adversarially, exactly as for a colleague’s pull request. The engineer’s output shifts from code to decisions: what to build, what to verify, what to reject.

Why does the harness matter more than the model?

Because the same model produces junior or senior work depending on what surrounds it: the context it is fed, the tools it may call, the tests that gate it, the conventions it inherits. That surrounding machinery is the harness, and building it well is harness engineering. We wrote up our own version of this lesson in harness engineering: the model is not the product. Teams that invest in the harness compound; teams that only prompt harder plateau.

What should a team adopt first?

The cheap discipline: a repository the agent can test itself in, a written convention file the agent reads, and review rules that treat agent code like any other author’s. Measure outcomes the way agentic coding maturity is actually measured: cycle time and defect rate, not lines generated. Start where failure is cheap, keep the verification gates human, and widen the agent’s scope as its track record earns it.

This entry was drafted with AI assistance.

Frequently asked questions

What is the difference between agentic engineering and vibe coding?

Verification. Vibe coding accepts AI-generated code largely unreviewed, which is fine for prototypes. Agentic engineering keeps specs, tests and review in the loop, which is what production requires.

Is agentic engineering a job title?

Mostly a practice, increasingly a role. Teams describe engineers who spend their day directing, reviewing and tooling coding agents rather than typing most of the code themselves.

Summarize this page with

See this working in a system we built