When an agent delegates to a sub-agent, the sub-agent runs in an isolated context: its intermediate results live only in its own memory and are discarded on return. A follow-up delegation that needs the same data has to re-fetch it from source. A conversation-scoped artifact scratchpad eliminates that redundancy by propagating a shared store to the orchestrator and every sub-agent it spawns.
How the scratchpad works
- Artifact capture. A tool result flagged as reusable is captured into the scratchpad automatically when the agent loop processes it — no extra round-trip to the language model. Entries are bounded by a configurable count ceiling.
- Shared by reference. The same store instance is propagated to every sub-agent context. A sub-agent receives a key index at delegation time so it can read prior artifacts with
scratchpad_readwithout knowing whether the orchestrator or a sibling produced them. - Explicit pinning. An agent can write a named entry with
scratchpad_writeto make it permanently visible to the rest of the session; auto-captured entries are managed by the weigher, pinned entries are not evicted.
Why it matters for multi-agent deployments
- Fewer redundant tool calls. Data that one sub-agent fetched — a query result, a resolved lookup, a computed summary — is available to every subsequent delegation without hitting the source again.
- Bounded memory. The store is Caffeine-backed with a configurable entry limit, so a long-running session does not accumulate unbounded state.
The scratchpad is opt-in per agent configuration and integrates with the existing delegation model. Agents that do not load the scratchpad toolkit continue to operate unchanged.