Agno v3.1.4 adds sandboxed Claude and Codex agents

Agno published v3.1.4 on October 16, 2026. The update moves the complete Claude or Codex harness into a per-session sandbox rather than running it on the application host.

What changed

• Two sandbox providers: ClaudeAgent and CodexAgent accept a sandbox parameter using customer-operated DockerSandbox or Amazon Bedrock AgentCoreSandbox.
• Recoverable execution state: checkpoints preserve the workspace and native conversation files. Failed or cancelled turns retain the previous recovery point; a missing native conversation stops restoration before model execution.
• Warm sessions: the default retention window is 300 seconds. Docker pauses the computer; AgentCore leaves it idle. A checkpoint is still captured after each completed turn.
• AgentOS integration: sandbox runs use the existing durable queue, reconnect and cancellation routes. Postgres coordination is required, and queue workers cannot use the default host-local checkpoint store.

These are the maintainers' release claims, not an independent sandbox security or performance assessment.

Where it fits

Agno's service guide describes AgentOS as the layer around a harness: it provides orchestration, authentication, streaming, session storage and background execution on infrastructure the operator controls. The application remains responsible for users, permissions and business operations. The guide recommends idempotent state-changing tools and testing multi-replica recovery with shared storage.

License and cost

The license included in this release is Apache-2.0. That is the repository's software license, not a promise of free model inference or sandbox infrastructure. No prices for those services are quoted in the release notes.

See the Agno hub, frameworks category, and earlier v3.1.3 background-run update.