Summary

The spec defines a third execution loop (Loop 3: Hybrid) that combines planning with per-step mini-ReAct loops and checkpointing. ReAct (Loop 1) and Plan-and-Execute (Loop 2) are implemented; Hybrid is not.

Design Spec Reference

• §6.5 Agent Execution Loop — Loop 3: Hybrid Plan + ReAct Steps

Scope

• High-level plan generation (3–7 steps)
• Per-step mini-ReAct loop with own turn limit
• Checkpoint after each step (progress summary)
• Optional replanning of remaining steps
• ExecutionLoop protocol compliance

Research Context

StructuredAgent AND/OR Trees (arXiv:2603.05294) — Consider as Loop 4 candidate:

• AND nodes (all children succeed), OR nodes (any child suffices), DFS traversal with 3 states
• 52.6% on WebArena vs 46.4% flat ReAct — significant improvement on long-horizon tasks
• Node repair: modify/add children within revision budget before cascading failure up
• Observation summarizer: compress full history into progress + constraint summaries
• Auto-select for complex/epic tasks via get_loop_type() -> "andor_tree"
• No code released — adapt algorithms

See research/structuredagent-andor-trees.md in project memory.