Agent-Driven Development¶

The way I build software in 2026.

Core principles¶

Code is free. Generation is no longer the bottleneck.
Humans define intent, not syntax. The expensive work is specs and decisions.
Evaluation pipelines are the gatekeeper, not manual human review. Evals and CI guard quality.

Workflow¶

Concept → Plan → Spec → Code → Feature — a hard rule. Every non-trivial change starts as a written plan or spec; agents work against the document, not against a chat thread.

Multi-agent setup¶

Specialised roles run in parallel or sequence:

Planner — turns concepts into specs
Developer — writes code against the spec
Tester — generates and runs tests
Code reviewer — checks quality, style, intent fit
Security reviewer — checks for vulnerabilities, secrets, anti-patterns
Ops / deploy — manifests, CI, rollouts

Orchestration patterns: orchestrator + sub-agents, parallel sub-agents, planner-executor, reviewer / critic.

Project rails¶

CLAUDE.md and AGENTS.md define design principles, commit conventions, and no-go zones
Spec / Plan / Feature registry (S0NN, P0NN, #FNNN) is referenced from code and PRs
IaC-as-truth — agents edit manifests in repo, not via ad-hoc CLI

Outcomes¶

pois.world — six microservices, seven shared libs, ~71k LOC, built solo agent-driven from day one (see pois.world — AI-driven build of an open map platform)
Production AI agents shipped at DesignWeltDeko with LLM/RAG evaluation in production
Bare-metal Kubernetes clusters at Google Distributed Cloud Hosted operated agent-driven via runbooks