💡 Three Patterns Worth Adopting Now That Mid-Conversation System Messages Are GA

With mid-conversation system messages now generally available (see this week's earlier entry), a few concrete patterns are worth building into long-running sessions:

• Phase transitions in multi-stage agents — an orchestrator moving from "research" to "implementation" phase can update the system message to reflect the new phase's constraints, rather than relying on the model to infer the transition from conversation flow alone.
• Permission tightening mid-session — if a session starts broad and narrows in scope (e.g. moving from exploratory research to a specific destructive operation), inject a system message reinforcing the narrower constraint at that exact point, rather than trusting an early-session instruction is still salient.
• Context injected from external state changes — if something relevant changes outside the conversation (a deploy completes, a ticket status updates), a system message is a cleaner way to surface that than working it awkwardly into the next user turn.