Framework Updates

Agno v3.0.0 overhauls database, adds offloading, and durable execution.

Agno v3.0.0 is a major breaking release fundamentally enhancing scalability and production readiness. The core database schema for agent runs has been rewritten, shifting from O(N²) to O(N) write amplification and removing item-size limits, which requires a database migration. New features like tool result and media offloading significantly reduce message sizes and associated costs, improving performance for large outputs. Additionally, the release introduces durable background execution for reliable, fault-tolerant agent operations, alongside a new CodeMode for advanced interactive programming with agents.

Mastra enables caller-driven experiments and workflow concurrency safety.

This release introduces caller-driven experiment APIs, allowing external orchestrators like Temporal to manage experiment execution while Mastra maintains the system of record. This provides critical flexibility for integrating evaluation pipelines into existing infrastructure. Workflow `resume()` calls are now atomically safe, preventing concurrent execution issues and enhancing system reliability in distributed environments. Engineers also gain granular control over server shutdown with new graceful drain options and can perform multi-turn LLM judge evaluations.

PydanticAI supports Anthropic 1.0.0 and its `httpx2` client requirement.

PydanticAI v2.33.0 introduces official support for the `anthropic` 1.0.0 SDK, which has migrated to `httpx2`. This update resolves runtime failures experienced with previous PydanticAI versions when `anthropic` 1.0.0 was unpinned or freshly installed. Engineers leveraging Anthropic models with PydanticAI must upgrade to v2.33.0 for compatibility. Additionally, any custom `http_client` passed to `AnthropicProvider` must now be an `httpx2.AsyncClient`.

PydanticAI enhances FunctionModel to accept any callable function.

PydanticAI's `FunctionModel` now accepts any Python callable for its `function` and `stream_function` arguments. This enhancement significantly increases flexibility for engineers defining tools and agent capabilities. It simplifies integration with existing codebases by removing previous constraints on function signatures. This is an additive improvement, requiring no migration.

CrewAI adds declarative conversational flows for agent interactions.

CrewAI introduces declarative conversational flows, providing a more structured way to define agent interactions. This allows engineers to synthesize built-in conversational methods, enabling declarations to directly drive the agent's conversational mode. The change aims to simplify the design and management of complex multi-turn conversations, leading to more predictable and maintainable agent behavior. Developers should explore the new declarative patterns to streamline agent development, with no explicit breaking changes noted.

LlamaIndex AG-UI now supports multimodal user input.

LlamaIndex's Agent-UI (AG-UI) now supports multimodal user input, allowing agents to process images, audio, video, and documents directly from the UI. This significantly expands the scope of interactive agent applications. Core framework enhancements include `Memory` components now accepting any `AsyncDBChatStore` for flexible custom backends and async support for `LLMRerank`. Additionally, this release adds support for new LLMs, including Claude Sonnet 5/Opus 5, GPT-5.6, and sets Gemini 3.7 Flash as the default for `google-genai` users, requiring review for potential impacts.

LangGraph adds support for decrypting replacement results.

This release introduces a new `decrypt replacement result` feature in the LangGraph SDK. This functionality is significant for engineers dealing with sensitive information within agent workflows, where data might be transformed or replaced and requires secure decryption. It enhances the ability to manage data privacy and security for intermediate or final outputs in production AI agent systems. Specific implementation details will likely be covered in the updated framework documentation.

Mastra adds durable execution and Cloudflare Sandbox for agents.

Mastra introduces durable execution for its Agents API, enabling long-running agents to persist state and recover without deploying new code, enhancing production reliability. This release also adds a `@mastra/cloudflare-sandbox` provider, offering a new remote workspace runtime option for agent execution and expanded deployment flexibility. For RAG systems, `GraphRAG` now supports `serialize()` and `deserialize()` to significantly reduce knowledge graph startup times. Additionally, sandboxes gain checkpointing capabilities for faster warm-starts, and the MCP protocol is upgraded for improved interoperability.

OpenAI Agents SDK enforces provider config changes, improves runtime security.

This release introduces a breaking change for `OpenAIProvider` configurations: if an explicit `openai_client` is used, `organization` or `project` parameters must now be specified directly on the `AsyncOpenAI` client instance. This tightens the provider contract, preventing conflicting options and ensuring correct behavior. Additionally, the SDK enhances runtime security by redacting sensitive function-tool outputs from replayable state, and improves operational visibility with isolated usage accounting for `RunState` checkpoints. Engineers should update their provider instantiation to avoid runtime errors.

PydanticAI updates instrumentation for tool results, supports xAI attachments.

PydanticAI v2.32.0 introduces updated instrumentation (v6) that now emits tool results under a `role: 'tool'` within the event stream. This change aligns with common LLM API response structures, simplifying integration for engineers building observation systems or custom UIs around agent execution flows. The release also adds support for the xAI attachment search lifecycle and surfaces OpenRouter web-search sources in `provider_details["annotations"]`, expanding data integration capabilities. Engineers consuming instrumentation events should update their parsing logic to accommodate the new `role: 'tool'` structure for tool result reporting.