Agent Guardrails and Controls: Applying the CORS Model to Agents ◆ goose
To handle these threats we propose removing all tool-call responses from the context window in-between user turns. This significantly increases the difficulty of performing “inter-turn” manipulation at the cost of occasionally forcing it to re-run tool-calls if it requires more precise historical values.
This is new (to me). The reduction of tool outputs on the next user turn.. is this feasible? i mean, it kills state cache so its very costly.
