Architecture AI Updates: November 11, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Kiran Prakash Describes Durable Execution as the Recovery Pattern for Agents That Run for Days or Weeks
Kiran Prakash published an article on martinfowler.com on November 10 arguing that long-horizon agents should run on durable execution, where the runtime records the input and output of every step that touches the outside world (model calls, ERP calls, supplier messages) in a log. After a crash the program reruns from the top, completed steps return their recorded results, and local state is rebuilt from the log; code between steps therefore runs again and must behave the same way on each replay. The worked example is a stock-replenishment agent that gets quotes, waits for a manager’s price approval, places an order and tracks delivery over weeks, using one long-lived entity workflow per SKU that starts a short process workflow per replenishment, with waits suspended on logged timers and messages so no process stays resident. Because retries give at-least-once behavior, the harness must pass a stable idempotency key such as a purchase order number, and compensating tools such as cancel_order undo effects found to be wrong. The article names Temporal, Cloudflare Workflows, Restate, Inngest and DBOS, and lists caveats: the log records what the agent saw rather than what happened, decisions can go stale during long waits, code changes while workflows are suspended can break replay, replay cost grows with log length, and durability does not protect against prompt injection from untrusted text such as supplier emails. Source