FROM VIBE CODING TO AGENTIC ENGINEERING: TEST-FIRST ORCHESTRATION
Engineering teams are shifting from vibe coding to disciplined agentic engineering that treats AI as test-driven collaborators and demands spec-first oversight....
Engineering teams are shifting from vibe coding to disciplined agentic engineering that treats AI as test-driven collaborators and demands spec-first oversight.
In a concise critique of “prompt DJ” development, Roger Wong summarizes Addy Osmani’s call for agentic engineering—engineers orchestrate coding agents, act as architects and reviewers, and enforce spec-first discipline instead of accepting whatever the model returns.
Simon Willison’s “First run the tests” pattern operationalizes this by making a test suite the entry point for any agent, turning TDD into a four‑word prompt and letting agents learn a codebase through its tests.
Hands-on workflows show how to scale this in practice, from a complete greenfield agentic setup to advanced agent teams comparing Claude Code and Codex, while case studies like DumbQuestion.ai underline the need for structured backlogs and cost-aware multi‑model choices.
Spec-first, test-led agent orchestration cuts rework and reduces regressions from AI-generated changes.
Without rigor and mentorship, ‘vibe coding’ erodes code quality and weakens the talent pipeline.
-
terminal
Bake “First run the tests” into agent prompts and verify agents can execute the suite locally and in CI before proposing diffs.
-
terminal
Evaluate multi-model agent routing for code tasks (quality, latency, cost) against your repos and enforce guardrails for token/JWT handling.
Legacy codebase integration strategies...
- 01.
Expose a standard test command in each service, enable agents to run it in CI, and require human review until test coverage stabilizes.
- 02.
Start with low-risk services, audit agent JWT/token flows, and pair juniors with seniors to review agent output and error patterns.
Fresh architecture paradigms...
- 01.
Define architecture and backlog first, scaffold tests up front, and codify a minimal agent playbook (spec, tests, implement, verify).
- 02.
Design least-privilege agent auth with JWT best practices and instrument cost/latency metrics to guide model selection.