The minimal governance stack

Here is the smallest set of constraints I have seen that changes outcomes materially: Roles, not vibes. Define narrow agent roles with hard boundaries.

    Example of a specialized "CodeFactory" configuration:

 * Architect: Defines ADRs (Architecture Decision Records) and contracts. Prohibited from writing implementation code.
 * API Generator: Implements backend endpoints strictly from OpenAPI specs. No business logic in the API layer; validation is generated from the contract.
 * Worker Generator: Implements asynchronous processing based on message contracts. Focus: idempotency and statelessness.
 * Frontend Generator: Creates UI components based on UX specs and API contracts. Mandatory handling of three states: loading, error, and empty.
 * DS (Data Scientist): Constructs ML pipelines and trains models. Abstracting complexity into dedicated modules/clients.
 * DevOps: Manages infrastructure, deployment scripts, and rollback playbooks. Focus: idempotency of operations and post-deploy health checks.
 * Critic: The automated Quality Gate. Prohibited from "fixing" code. Only issues verdicts (Pass/Fail) based on linters, tests, and contract parity.

    Contracts-first

Before code, freeze interfaces: OpenAPI for HTTP, message contracts for async events, DB schema and invariants. This prevents "LLM-driven interface drift" where the agent silently changes request shapes because it feels nice. * Read-evidence, not trust - Ban "I assume this file contains…" behavior. Require file reads before edits and require the agent to cite what it saw: which functions exist, which types exist, where the integration points are. It is about forcing contact with reality.\ * Determinism over cleverness - Prefer minimal diffs, explicit types, explicit invariants, explicit error paths, idempotent workers, no "magic" implicit behavior. LLMs love cleverness because cleverness sounds correct. Determinism survives maintenance. * Hard gates at the boundary - A "critic" role runs linters, unit tests, integration tests, contract validation, migration checks, deploy health checks. If it fails, the pipeline stops. Not "warns". Stops. * Explicit handoffs - Every step ends with: what changed, what assumptions were made, what is now the source of truth, who owns the next step.

Code is cheap now, but engineering is not. And developers are learning now that code and engineering are not the same thing.