LLM Wiki — Synaptic Lattice

Cooldown Heuristics

Properties1
tagsprompt-engineering, workflow-pattern

Leitura de 2 min

Cooldown Heuristics

Definition (confidence: high)

A workflow pattern where the tool detects rapid repeated failures or loops and backs off — pausing, summarizing the situation, and asking the user for guidance. Prevalence: LOW (~8%, ~4 tools: Claude Code, Cline).

Mechanics (confidence: high)

  1. Track error frequency across tool calls or actions
  2. If errors exceed a threshold within a time window, trigger cooldown
  3. Pause execution and summarize what went wrong
  4. Present the summary to the user and ask for guidance before proceeding
  5. Resume only after user input

Current Limitations (confidence: high)

  • Only Claude Code has explicit cooldown heuristics, and it monitors error frequency, not token pressure
  • No tool handles graceful degradation under token pressure — tools silently drop context, lose focus, or produce truncated output when the context window is nearly full
  • This is the only pattern in the corpus that addresses agentic loop stability

chain-of-thought-pattern — CoT is the most common workflow instruction (~60% of tools). Cooldown heuristics are a safety net when CoT-driven execution fails.

task-state-machine — The task lifecycle pattern (pending → active → done → verified) pairs naturally with cooldown heuristics. When verification fails repeatedly, cooldown triggers.

claude-code-system-prompt — Claude Code implements both cooldown heuristics and the task state machine, making it the most robust agentic workflow in the corpus.

cooldown-vs-decision-tree — A comparison of heuristic cooldown vs deterministic decision tree error recovery approaches.

Implications (confidence: high)

Cooldown heuristics are an essential but under-implemented safety mechanism for autonomous agents. The biggest gap is token pressure degradation — no tool handles the case where the context window is nearly full. As agents become more autonomous and handle longer sessions, graceful degradation under resource pressure will become critical. The cooldown pattern for errors is a good foundation, but it needs to extend to resource constraints.

Sources

  • [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]

Visão de gráfico

Backlinks