Harness engineering for coding agents uses layered configuration surfaces¶
Insight: Effective harness engineering uses distinct configuration layers: CLAUDE.md/AGENTS.md (kept under 60 lines), MCP servers (selective, context-aware loading), skills (progressive disclosure), sub-agents (context firewalls), hooks (lifecycle automation), and back-pressure mechanisms (context-efficient error reporting). The philosophy prioritizes reactive, shipping-focused iteration over preemptive optimization.
Detail: Core principle: 'When the agent fails, take the time to engineer a solution so it doesn't fail that way again.' This emphasizes post-hoc harness improvements over trying to anticipate all failure modes upfront. Sub-agents create context isolation by moving discrete tasks to separate sessions.