agent-loopslisted

# Agent Loops Enable LLMs to reason, plan, and take autonomous actions. ## Key Decisions | Parameter | Recommended | Notes | |-----------|-------------|-------| | Max steps | 5-15 | Higher for research, lower for actions with side effects | | Temperature | 0.3-0.7 | Lower for deterministic tasks, higher for creative | | Memory window | 10-20 messages | Summarize older context to avoid token overflow | | Validation frequency | Every 3-5 steps | Self-check to prevent drift | ## Patterns Choose based on task type: - **ReAct** (Thought/Action/Observation loop) - Default for most tasks - **Plan-and-Execute** - For multi-step tasks where upfront planning reduces errors - **Self-Correction** - When output quality must be validated before returning See [agent-workflow-template.ts](${CLAUDE_SKILL_DIR}/templates/agent-workflow-template.ts) for a complete TypeScript implementation. ### Karpathy Loop (Autonomous Iteration) Metric-driven autonomous improvement loop. Establish baseline, then iterate: hypothesize, modify, evaluate, keep if improved, discard if regressed. Continue until budget exhausted or goal reached. **Components**: mutable target + measurable metric + fixed budget + binary keep/discard **Use cases**: performance optimization, coverage improvement, ML training, cost reduction **Key constraint**: single metric with clear direction (minimize/maximize) removes ambiguity **Safety**: git-based rollback, file allowlists, correctness gates, stuck detection See `/experi