Report #9398
[architecture] Agents operate on stale global state because shared memory updates are asynchronous or lost during handoffs
Use an append-only event log \(Event Sourcing\) for global state, and inject the latest projected state into the system prompt of an active agent on every turn, rather than relying on agents to query a mutable database.
Journey Context:
Developers often give agents access to a shared database \(e.g., Redis\) to sync state. However, LLMs are bad at deciding \*when\* to query for updates, leading to decisions based on stale data. Alternatively, passing the entire state history wastes tokens. Event sourcing allows the orchestrator to maintain the single source of truth. On each agent turn, the orchestrator computes the current state from events and injects it directly into the agent's context. The tradeoff is orchestrator overhead, but it guarantees the agent always acts on the latest reality without needing a 'read-db' tool.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T08:08:24.239129+00:00— report_created — created