Summary

Add stagnation detection to execution loops (ReactLoop, PlanExecuteLoop) that identifies when an agent is repeating the same tool calls without making progress, and intervenes.

Motivation

From MADQA paper (arXiv:2603.12180, Snowflake/Oxford): agents persist in unproductive loops, achieving ~20% below oracle accuracy via brute-force exhaustive search rather than strategic reasoning. Current execution loops only have max_turns ceiling and budget_checker as cost gate — no mechanism to detect repetitive patterns.

Design

• Analyze TurnRecord.tool_calls_made history across recent N turns
• Detect repetitive patterns: same tool + same args repeated, same documents accessed, no new information gained
• On detection, two responses:
  1. Corrective prompt injection: "You appear to be repeating the same actions without progress — try a different approach"
  2. Early termination: Terminate with STAGNATION termination reason (new enum value)
• Configurable sensitivity: window size (N turns), similarity threshold, max repeated patterns

Affected Files

• src/ai_company/engine/react_loop.py
• src/ai_company/engine/plan_execute_loop.py
• src/ai_company/engine/loop_protocol.py (add STAGNATION to termination reasons)
• src/ai_company/engine/run_result.py

Research

• MADQA: Strategic Nav vs Stochastic Search (arXiv:2603.12180) — primary motivation
• Connects to Memex (arXiv:2603.04257) redundant tool call penalty concept
• Raw data already available in TurnRecord.tool_calls_made