What Shipped in 1.14.3

crewAI 1.14.3 delivers across four areas: new lifecycle and checkpoint capabilities, expanded cloud and sandbox integrations, a measurable cold-start performance gain, and a set of bug fixes that stabilize serialization and replay behavior [1]. The release also patches two security dependencies, making it a recommended update for teams running crewAI in production. Contributors include ten individuals spanning core maintainers and external collaborators [1].

Checkpoint and Fork Capabilities

The most operationally significant additions in this release center on checkpointing. The framework now emits lifecycle events for checkpoint operations, giving operators visibility into when checkpoint state is written and resumed. For teams building observability pipelines or audit trails around long-running agent workflows, these events provide hooks that were previously absent [1].

Standalone agents, which run outside a full crew context, gain checkpoint and fork support in this release. Fork support means an agent’s execution state can be branched, allowing operators to resume from a saved point along a different path without rerunning the entire workflow from scratch. A companion fix ensures that task_started events are emitted correctly on fork resume, and the checkpoint terminal user interface has been redesigned alongside that change [1]. A separate fix resolves replay of recorded method events on checkpoint resume, addressing a gap where replayed events were not being correctly reconstructed after a checkpoint load [1].

Bedrock V4 and Sandbox Integrations

Amazon Bedrock V4 support is included in this release, extending compatibility for teams that use AWS-hosted models as the inference backend for their agents [1]. No migration steps beyond the version upgrade are described in the release notes.

Two sandbox integrations also ship in 1.14.3. Support for e2b provides a code execution sandbox environment, and Daytona sandbox tools are added for what the release notes describe as “enhanced functionality” [1]. These additions expand the set of tool environments available to agent developers without requiring external patching.

On the Azure side, the release adds a fallback to DefaultAzureCredential when no API key is provided, reducing configuration friction for operators using managed identity or other credential chain approaches in Azure deployments [1].

Cold-Start Optimization and Security Patches

A performance refactor targeting the MCP SDK and event-type handling reduces cold-start time by approximately 29 percent [1]. Cold-start latency affects how quickly an agent process becomes ready to handle its first task, making this optimization relevant for deployments that frequently spin up new agent instances, such as serverless or on-demand execution patterns. The release notes attribute the gain specifically to optimizations in MCP SDK initialization and event-type processing, without detailing further implementation specifics [1].

Two security dependency bumps accompany the performance work. The lxml library is now pinned to version 6.1.0 or higher, and python-dotenv is bumped to version 1.2.2 or higher, both in response to security patches in those upstream packages [1]. Operators who manage their own dependency pins should verify that their environment satisfies these new lower bounds.

Bug Fixes Affecting Serialization and Replay

Several fixes in 1.14.3 address serialization correctness, which is critical for checkpoint reliability. The serialization of initial_state class references as JSON schema is corrected, and serialization of Task class-reference fields used during checkpointing is also fixed [1]. Together, these changes reduce the risk of corrupted or unresumable checkpoint state.

Additional fixes cover a range of execution edge cases. The execution_id field is now kept separate from state.id, resolving an identity collision that could affect state tracking across runs. Metadata-only agent skills are preserved rather than dropped, implicit @CrewBase names are propagated correctly to crew events, and execution metadata is merged properly on duplicate batch initialization [1].

On the model compatibility side, a fix preserves thought_signature in Gemini streaming tool calls, addressing a regression that affected agents using Gemini models with streaming enabled. A fix for BaseModel results in the guardrail retry loop rounds out the stability improvements [1].

FAQ

Q. Does upgrading to 1.14.3 require changes to existing checkpoint implementations? The release notes do not describe breaking changes to the checkpoint API. The new lifecycle events and the standalone-agent fork support are additive, but operators using checkpoints should verify that the serialization fixes for initial_state and Task class-reference fields do not alter previously saved checkpoint formats [1].

Q. What is the minimum version of lxml and python-dotenv required after this upgrade? After upgrading to 1.14.3, lxml must be at version 6.1.0 or higher and python-dotenv must be at version 1.2.2 or higher, as specified in the security patch entries for this release [1].

Q. Does the 29 percent cold-start reduction apply to all deployment configurations? The release notes state the reduction comes from MCP SDK and event-type optimizations and describe it as approximately 29 percent, but do not specify whether the improvement is uniform across all deployment types or model backends [1].

Q. Is Amazon Bedrock V4 support a breaking change for existing Bedrock users? The release notes list Bedrock V4 support as a new feature without describing deprecation of prior Bedrock versions or required migration steps [1].

Q. What does fork support for standalone agents enable that was not possible before? Fork support allows an agent’s execution state to be branched from a checkpoint, enabling resumption along a different path without rerunning the full workflow. Prior to this release, standalone agents did not have checkpoint or fork capabilities at all [1].

Key takeaways

  • crewAI 1.14.3 adds lifecycle events for checkpoint operations and introduces checkpoint and fork support for standalone agents, enabling state branching and improved observability in long-running workflows [1].
  • Amazon Bedrock V4 support, e2b sandbox integration, and Daytona sandbox tools expand the set of inference backends and execution environments available to operators [1].
  • MCP SDK and event-type optimizations reduce cold-start time by approximately 29 percent, benefiting deployments that frequently initialize new agent instances [1].
  • Security patches bump lxml to 6.1.0 or higher and python-dotenv to 1.2.2 or higher; operators with pinned dependencies should update their environment constraints [1].
  • Multiple serialization and replay fixes improve checkpoint reliability, including corrections to initial_state JSON schema serialization and Task class-reference field handling [1].