Framework Updates

PydanticAI adds Bedrock native JSON output and strict tool calls.

PydanticAI now supports native JSON output and strict tool calls for AWS Bedrock models. This enhancement improves reliability and control when interacting with Bedrock-hosted LLMs, enabling more predictable function calling and structured data extraction crucial for robust AI agent workflows. Additionally, this release introduces several deprecations ahead of v2, impacting `gateway/gemini:` prefixes, `StreamedResponse.usage()`, and evaluation class construction. Engineers should review these deprecations to prepare for future breaking changes and align with upcoming API standards.

PydanticAI adds `gemini-3.5-flash` model support.

PydanticAI now includes support for the `gemini-3.5-flash` model. This integration expands the available model choices for engineers building AI agents, particularly those within the Google Gemini ecosystem. The `gemini-3.5-flash` model is optimized for speed and cost-efficiency, making it a suitable option for high-throughput applications or latency-sensitive tasks where a lighter, faster model is advantageous. Engineers can begin using this new model by specifying it during model instantiation, leveraging the existing PydanticAI abstraction for model interactions.

Agno integrates Google Antigravity and Gemini Managed Agents.

This release significantly expands Agno's external agent capabilities by integrating Google's Antigravity API and providing first-class support for Gemini Managed Agents like Deep Research and Antigravity sandboxes. Engineers can now leverage these powerful, managed services directly through dedicated Agno agents or toolkits for advanced tasks, including autonomous research with citations. The release also delivers critical security hardening against path traversal vulnerabilities in filesystem-touching tools, improving overall production readiness, and bumps the minimum Python version to 3.9. Additionally, major bug fixes restore streaming event reliability for AG-UI Teams and ensure correct multi-turn interaction statefulness for Anthropic and Gemini-based agents.

PydanticAI adds OpenAI response token counting; refines agent retry configuration.

PydanticAI v1.98.0 introduces `OpenAIResponsesModel.count_tokens` for precise input token measurement on OpenAI API responses. This feature allows engineers to accurately track costs and manage prompt usage for agent interactions, critical for optimizing production deployments. Additionally, the `Agent` API has been updated, replacing `tool_retries` and `output_retries` with a consolidated `retries: int | AgentRetries` parameter. This change streamlines agent configuration for failure handling and requires updates for existing agent definitions using the previous retry parameters.

CrewAI deprecates `CrewAgentExecutor` and introduces state restoration.

This CrewAI release defaults agent execution to `AgentExecutor`, deprecating `CrewAgentExecutor` to standardize the agent execution model. This change may simplify integration and future agent development patterns. A new `restore_from_state_id` parameter enables resuming agent runs from a specific state, enhancing workflow resilience and debugging capabilities. Additionally, the CLI has been extracted into a standalone `crewai-cli` package, requiring separate installation for command-line interactions.

Mastra adds Fine-Grained Authorization (FGA) route policy controls.

Mastra now includes Fine-Grained Authorization (FGA) route policy controls, enabling developers to centrally define and enforce access rules for protected routes. This new capability, with built-in resource resolution and resolver hooks, streamlines security implementation for production AI agent applications by mapping routes to specific resource types, IDs, and permissions. The release also introduces a `favorites` storage domain for agents and skills, allowing users to mark and filter them by visibility and favorite status, with support across various storage backends. Engineers will also benefit from improved token estimation in `@mastra/memory` for large file parts, enhancing the accuracy of observational memory and preventing excessive history replays.

Mastra adds ACP coding agents and xAI Grok Voice integration.

This release introduces the new `@mastra/acp` package, enabling integration of ACP-compatible coding agents as tools or subagents within Mastra, complete with incremental streaming support. It also adds a real-time voice integration for xAI's Grok Voice API via `@mastra/voice-xai-realtime`, allowing agents to handle live audio interactions. Additionally, agents can now include `metadata` for enhanced filtering and dynamic resolution, improving agent management. The `workspace.read_file` tool has been upgraded to better handle various file types, providing native model parts for media and smarter binary reporting, while new enterprise foundations lay groundwork for admin policies and RBAC.

PydanticAI revises Google providers and deprecates `fasta2a` integration.

PydanticAI v1.97.0 significantly revises integration with external services and stabilizes core APIs. Google provider configuration is updated, splitting `GoogleProvider` and `GoogleCloudProvider` with new IDs, requiring changes to existing deployments. The bundled `fasta2a` integration is deprecated; users must now integrate via `fasta2a.pydantic_ai` (requiring `fasta2a` v0.6.1) as an external dependency. Additionally, the `pydantic_graph` API moves out of beta, and the `OnlineEvaluator` gains the ability to run evaluations on failed calls, enhancing debugging capabilities.

Agno adds Gemini stateful interactions and user data isolation.

Agno now supports Google's new stateful interactions API through a dedicated `GeminiInteractions` model, enabling more persistent and context-aware conversational flows with Gemini. This allows engineers to build more sophisticated multi-turn agents by leveraging built-in state management. A new opt-in per-user data isolation layer for AgentOS authenticated endpoints significantly improves security and privacy for multi-tenant applications. Additionally, `allowed_hosts` can now restrict outbound fetches in knowledge readers, mitigating SSRF vulnerabilities.

PydanticAI deprecates `Agent` constructor parameters for upcoming V2.

PydanticAI has deprecated specific `Agent` constructor parameters, including `prepare_tools`, `prepare_output_tools`, and `event_stream_handler`. This change is a significant part of the V2 preparation, signaling a move towards a more explicit capability-based agent configuration. Engineers currently using these parameters should begin migrating to the new `PrepareTools`, `PrepareOutputTools`, and `ProcessEventStream` capabilities. Adopting these new patterns now will streamline the transition to the upcoming PydanticAI V2 release, which is planned for June.