Journal Then Replay

Durable-execution engines persist step results in a journal or event history and recover a crashed agent run by replaying it, with deterministic step identifiers acting as memoisation keys so work resumes from the failure point.

Evidence: strong

Agents are stateful and their errors compound, so restarting a long run from scratch after a crash wastes work and can repeat side effects. Anthropic's multi-agent research system resumes from the failure point rather than restarting, and the wider durable-execution industry has converged on one way of doing that.

The shared model has three parts: orchestration code that is deterministic, side effects wrapped in journaled "steps", "activities" or "tasks" whose results are memoised, and recovery by replaying the journal. Temporal runs the agent loop as a workflow, with model and tool calls as retried activities and a full event history to replay. Restate records each ctx.run() step in a journal. DBOS checkpoints workflows and steps to Postgres without an external orchestrator. Inngest memoises step.run results so a restart replays from the last successful step. LangGraph checkpointers save per-thread state snapshots, though its in-memory saver does not survive restarts and checkpoints can grow without bound. Cloudflare's rules spell out the discipline this demands: step names are the cache key and must be deterministic, only step return values persist, and logic outside steps may run more than once, so calls must be idempotent.

Pi Durable is the notable variant. It does not deterministically replay user code; it persists explicit task phases and checkpoints, closer to a persisted state machine, so code need not be deterministic between checkpoints. Its phase discipline is described in the effect sandwich, and its parent-child structure in task ownership trees.

What replay does not give is exactly-once execution of the effect itself, a contested point covered in exactly-once only after journaling. A journaled engine is a reasonable substrate for a multi-step agent run, provided that every step's key is stable, every external action is idempotent or checked before it is repeated, and human waits are modelled as durable approval state. File-based externalised progress artefacts offer a lighter alternative where full replay is not needed.

Sources

  • Temporal blog (Cornelia Davis), 10 Jul 2025 — https://temporal.io/blog/durable-execution-meets-ai-why-temporal-is-the-perfect-foundation-for-ai
  • Restate documentation — https://docs.restate.dev/ai/patterns/durable-agents
  • DBOS blog (Qian Li), 24 Feb 2025 — https://www.dbos.dev/blog/durable-execution-crashproof-ai-agents
  • Inngest blog (Charly Poly), 19 Feb 2026 — https://www.inngest.com/blog/durable-execution-key-to-harnessing-ai-agents
  • Cloudflare Workflows documentation — https://developers.cloudflare.com/workflows/build/rules-of-workflows/
  • LangGraph documentation — https://docs.langchain.com/oss/python/langgraph/durable-execution
  • Anthropic Engineering, 13 Jun 2025 — https://www.anthropic.com/engineering/multi-agent-research-system
  • Earendil Engineering, 1 Oct 2026 — https://earendil.com/posts/pi-durable/