Glossary term
Glossary term
Agentic Systems
Intermediate text space where a model writes out its working or reasoning before producing a final answer.
Wei et al. (2022) chain-of-thought prompting uses the response text itself as a scratchpad - the model writes 'Let me think step by step...' and uses the generated tokens as a working memory for multi-step problems.
OpenAI o1's extended thinking traces are a structured scratchpad - the model generates a long internal reasoning chain (shown as a collapsed 'thinking' block in the UI) before the final answer.
Anthropic's extended thinking mode for Claude 3.7 Sonnet uses a scratchpad mechanism - the model reasons in a separate 'thinking' block that does not count toward the output token limit.