Reinventing.AI

AI Agent Trends

AI Agents Trends: Five Open Source Agent Releases in One Week Add Crash Recovery and Change Upgrade Paths

OpenAI Agents SDK 0.25.0, Mastra 1.84.0, OpenClaw 2026.11.5, Agno 3.2 and goose 1.60.0 shipped between November 12 and 17. What changed and what breaks.

AI Agent Insights Team6 min
Open Source AI Agent Framework Releases

Five open source agent projects published releases between November 12 and November 17: the OpenAI Agents SDK, Mastra, OpenClaw, Agno and goose. Their release notes share two themes. Most of them add ways for an agent to continue after a crash or a restart. Most of them also contain changes that require action from anyone who upgrades.

For solo operators and small teams, the second theme matters as much as the first. These projects publish more than one release a week, and an upgrade that removes a provider or changes who can approve an action alters a working setup. The sections below draw on each project's release notes on GitHub.

OpenAI Agents SDK 0.25.0 removes a sandbox provider

The release notes for version 0.25.0 of the OpenAI Agents SDK for Python, dated November 17, open with a removal. The beta Runloop sandbox provider is gone. The notes tell Runloop users to preserve their workspace data before switching, to move to another supported provider such as Modal, or to stay on version 0.24.0 while they migrate.

A default also changed. The session class that compacts long conversations now uses the gpt-6-luna model unless a different one is set, and the notes give the setting that keeps the previous choice. An agent that relied on the old default will run its compaction on a different model after the upgrade.

The recovery additions are opt in. A handoff between agents can now recover from a failure when the developer supplies a failure function, and leaving it out keeps the earlier behavior, in which the error stops the run. Calls that fail or end incomplete keep the token usage the provider reported, so the cost of a failed run can be counted. The SDK is published under the MIT license.

Mastra 1.84.0 restarts crashed workflows

Mastra, a TypeScript agent framework whose core is licensed under Apache 2.0, published core version 1.84.0 with a workflow recovery loop. According to the notes, the loop restarts crashed runs on the default engine without waiting for another instance to boot, and a leader election lease coordinates the work across instances. On Redis or Valkey Streams and on Google Cloud Pub/Sub, instances no longer restart every run at boot. The message broker redelivers unacknowledged steps instead, which the notes say prevents the same step from running twice at once.

The release also adds an experimental storage format for agents and event driven workflows. Runs are stored as a header plus one row per step, so a long run writes only what changed. The notes give one example in which written data fell from 75 MB to 3.1 MB. The format is opt in and labeled experimental.

Four breaking changes are listed. The one most likely to affect stored data concerns the alpha Knowledge feature: scope matching rules changed, existing data is not rewritten, and the notes warn that some scopes can appear new or empty until their IDs are updated.

OpenClaw 2026.11.5 keeps queued messages across restarts

OpenClaw 2026.11.5, published November 17 under the MIT license, lists restart safety first among its highlights. Queued messages, send permissions and final chat writes are retained when the Gateway restarts. An update that fails at startup is rolled back. A configured primary account no longer blocks fallback to a second model route. Scheduled jobs are no longer aborted by a configuration reload, and late scheduler wakeups no longer drop jobs silently.

The breaking change applies to people who build on OpenClaw. Third party harnesses and plugins must change how they register active runs and how plugin handlers control sessions. The notes say bundled components are already migrated. Earlier coverage on this site goes deeper on replies that survive restarts and on update rollback.

Agno 3.2 moves to A2A v1.0 and narrows Slack approvals

Agno 3.2.0, published November 12 under Apache 2.0, upgrades the project to version 1.0 of the Agent2Agent protocol, which lets agents from different systems call each other. The notes tell users to upgrade the Agno server and client together. Clients on protocol version 0.3 keep working through a compatibility layer by default.

The Slack change affects daily use. The interface now runs on Slack's Bolt SDK, and only the person who started a run can approve or deny it, unless other people are named as approvers in the configuration. A team where one person starts agent runs and another approves them has to add that second person before upgrading. The release also makes Redis optional for sandboxed agents on the durable queue. Postgres and a shared checkpoint store now handle cancellation, resume and recovery across replicas.

goose 1.60.0 checks schedules before changing them

goose 1.60.0, published November 17 under Apache 2.0, is mostly a list of fixes. The agent now validates a cron expression before it changes an existing schedule. Request logs and temporary directories are kept private on shared systems. The release also drops the separate Skills extension in favor of a built in skill operation, which is the item to check for anyone who configured that extension by name.

What the five releases have in common

Four of the five add recovery from a restart or a crash as a feature of the framework itself: opt in handoff recovery in the OpenAI SDK, the recovery loop in Mastra, retained queues in OpenClaw and checkpoint based recovery in Agno. Recovery is becoming part of the framework instead of code each builder writes separately. The site's guide to recovering failed runs covers the parts that remain with the operator, such as deciding which steps are safe to repeat.

Three of the five also change behavior on upgrade in ways an operator may not notice at first: a different default model, a narrower approval rule, and a scope rule that can make stored data appear empty. This is the same class of problem described in yesterday's report on model retirements and agent reliability.

Practical steps

The following steps are general guidance drawn from the release notes above.

Pin the version. Each of these projects published more than one release in the past week. A fixed version number in the install file means an upgrade happens when the operator chooses.

Read the breaking changes section first. Four of the five releases label their breaking changes or removals, and the notes name the settings to change.

Check who can approve. After an upgrade that touches approvals, confirm that the right people can still approve and that nobody new can. The guide to permissions and review lists what to verify.

Test recovery on purpose. Stop the process in the middle of a run and observe what resumes, what repeats and what is lost.

Name the model explicitly. A default can change between versions. A named model changes only when the operator edits it. Readers new to the category can start with the agent fundamentals.

What to watch

The OpenAI notes say applications that share checkpoints across workers should upgrade the readers first. Mastra's new storage format is experimental and its Knowledge feature is in alpha, so both can change again.

Sources