Thoughtworks write-up on harness engineering for reliable agentic systems

Category: Agentic coding
Status: discussion
Sources: martinfowler.com, discussion
Summary: A practitioner article by Sarang Sanjay Kulkarni (Thoughtworks) published on martinfowler.com on 2026-06-16 describes the PRINCE production agentic system and argues reliability comes from context engineering (routing stage-specific context to each agent rather than one large prompt) and harness engineering (orchestration, tool boundaries, state persistence, retries, provider fallbacks, validation, reflection loops, observability, and human review). Concrete practices include three reflection types (process, data, draft), hybrid RAG plus Text-to-SQL retrieval, agent state persisted in PostgreSQL for failure recovery, automatic LLM provider fallback, and daily live-traffic evaluation with Langfuse. The post reports no quantitative metrics. The thread reached the HN front page (149 points).
Why it matters: It is a durable architecture write-up on building production agent workflows, though it offers patterns rather than measured results.