Treat AI software delivery as an orchestrated factory with explicit multi-lane roles, structured design debate, autonomous implementation, and synthesis-based review—rather than single-model coding sessions.
art-2026-03-02-001 · design-lane · confidence=medium · evidence=medium
Structured multi-round debate protocol before coding, with explicit challenge/concession behavior to force tradeoff surfacing.
source · research/kb/queue/queue-art-2026-03-02-001-agentic-software-factory-case-study.md
Enterprise agent systems should be designed as composable pattern stacks rather than single-agent chat interfaces; durable production value depends on combining tool use, reflection, planning, multi-agent orchestration, and adaptive reasoning.
art-2026-03-02-002 · design-lane · confidence=medium · evidence=medium
Five-pattern taxonomy: tool use, reflection, planning, multi-agent orchestration, and ReAct loops.
source · research/kb/queue/queue-art-2026-03-02-002-azure-agent-factory-patterns.md
Ralph is an autonomous coding loop runner that repeatedly executes fresh AI coding sessions against a PRD task list (passes: false -> true) until completion, with memory persisted outside the active context window.
art-2026-03-01-030 · agentic-software-factory · confidence=high · evidence=high
Fresh-instance execution discipline: each iteration starts with clean context, reducing drift from long conversational histories.
source · research/kb/queue/queue-art-2026-03-01-030-ralph.md