Framework Updates

PydanticAI optimizes OpenTelemetry serialization, avoiding O(n²) instrumentation cost.

PydanticAI v2.17.0 introduces a critical performance optimization for OpenTelemetry (OTel) by caching per-message serialization, effectively eliminating an `O(n²)` instrumentation cost. This change significantly reduces overhead for engineers utilizing OTel for monitoring AI agent systems, ensuring that observability does not degrade performance, especially with complex or high-volume interactions. Additionally, `RequestUsage` and `RunUsage` models now support arbitrary fields, future-proofing the framework for more detailed and custom `genai-prices` integration and operational cost tracking.

Agno adds Slack peer-agent communication and Scavio Google v2 API.

Agno v2.8.1 introduces a `respond_to_other_agents` flag for Slack, enabling direct peer-to-peer agent communication within channels. This feature is critical for developing sophisticated multi-agent systems requiring collaborative interaction beyond simple user-agent dialogues. Additionally, the release includes support for Marengo video embeddings via TwelveLabsTools, expanding agent multimedia processing capabilities. A breaking change for Scavio Google Search users is included, requiring an update to the v2 API for improved localization and paging.

PydanticAI enhances moderation, tool error handling, and adds new Gemini models.

This release introduces significant new features, including advanced moderation capabilities for OpenAI and Google Cloud (via Model Armor), providing more control over agent safety. A new `ToolFailed` primitive allows for explicit model-visible failures without automatic retries, enabling more nuanced error handling within agent logic. Additionally, PydanticAI now supports the latest `gemini-3.6-flash` and `gemini-3.5-flash-lite` models, expanding the available model ecosystem. Engineers should review the new moderation settings and consider integrating `ToolFailed` for improved agent robustness.

PydanticAI adds prompt caching and durable toolsets for agents.

This release introduces explicit prompt caching for OpenAI's `gpt-5.6` models, alongside enhanced durable execution that now supports `DynamicCapability` toolsets, integrating with DBOS steps and Prefect tasks. Prompt caching can significantly reduce LLM inference costs and latency for repetitive queries, while `DynamicCapability` enables more robust and stateful agent workflows within distributed environments. Engineers also gain granular control with the ability to override the tool-retry budget and expose OpenAI moderation results. Note the deprecation of Exa search common tools; migration to the new `ExaSearch` capability in Pydantic AI Harness is recommended for continued functionality.

PydanticAI introduces new durability capabilities, deprecates wrapper agents.

PydanticAI v2.14.0 introduces first-class `Temporal`, `DBOS`, and `PrefectDurability` capabilities, replacing previously deprecated wrapper agents. This provides a more integrated and robust approach for building resilient AI agents that require state persistence and fault tolerance. Additionally, a new `reasoning_effort` setting allows fine-grained control over Mistral model behavior within thinking settings. Engineers currently utilizing the older wrapper agents for durability should plan to migrate to these new dedicated primitives.

CrewAI adds authentication for skill registry downloads.

CrewAI now supports authentication for skill registry downloads. This feature allows engineers to secure access to private skill repositories, ensuring that only authorized agents or systems can retrieve skills. It significantly enhances the integrity and security posture of agent deployments, which is critical for production environments handling sensitive data or proprietary logic. Users can configure authentication credentials when specifying skill registry sources.

Agno v2.8.0 adds isolated agent environments and powerful new scorers.

Agno v2.8.0 significantly upgrades its evaluation suite by introducing `agno.environments` for isolated, K-shot evaluation rollouts and `agno.scorer` for converting agent runs into precise numeric scores. These new primitives empower engineers to robustly test agent performance, generate high-quality SFT data from successful attempts, and implement custom evaluation logic. Two critical breaking changes impact existing evaluations: `ReliabilityEval` now strictly matches tool executions, potentially altering previous pass/fail verdicts, and the hardened judge prompt may shift LLM-based scores and token counts. Users should re-run existing evaluations and review results post-upgrade to account for these changes, ensuring their agent assessments reflect these more rigorous criteria.

PydanticAI adds content filter error handling, cache metrics, and model hooks.

PydanticAI v2.13.0 introduces new capabilities for handling content filter errors, alongside `cache_hit_ratio` metrics for usage analysis. Engineers can now programmatically respond to content moderation flags using `RaiseContentFilterError` and gain deeper insights into agent performance and cost with cache hit ratios. Additionally, new `get_model`, `resolve_model_id`, and `for_agent` hooks provide enhanced flexibility for integrating custom models and managing agent interactions.

CrewAI promotes Skills Repository out of experimental status.

The CrewAI Skills Repository has been moved out of experimental status, signaling its stability and readiness for production environments. This change allows engineers to confidently build, share, and manage reusable agent capabilities with reduced risk of breaking changes or instability. It streamlines the development and scaling of sophisticated agent systems by standardizing skill management for production use.

Agno adds Superserve sandbox, Plivo tools, and observability integration.

Agno v2.7.4 introduces `SuperserveTools` for secure execution of agent-generated code within a Firecracker-based sandbox, crucial for robust, long-running agent deployments. New `PlivoTools` expand agent capabilities to include SMS and voice calls, enabling broader real-world interaction. Additionally, an observability integration with The Context Company provides essential tracing for production agent runs. Engineers should note the deprecation of `session_state` in Workflow Router selectors and Condition evaluators, favoring `run_context` for future-proofing.