The 5-Piece Kit: Replicable Architecture Skeleton

This page is a “skeleton view”: the five modules that tend to separate demo-grade agents from long-running systems.

The kit

1. Execution state machine: run / attempt / subscribe / active runs registry
2. Context engineering + recovery: guards, hygiene, pairing repairs, compaction retries, timeout snapshots
3. Tool safety: layered policy + human approvals for risky execution
4. Model fallback + error normalization: structured attempts and abort semantics
5. Subagents + skills snapshots: isolated runs, recovery, hot updates

See the corresponding deep-dive modules:

• Agent execution
• Context engineering
• Tool policy + exec approvals
• Model fallback + robustness
• Subagents + skills snapshots