Agent Teams

This guide teaches how to build a grade-and-revise loop using Outcomes in Claude Managed Agents, where a writer agent drafts a cited research brief and a stateless grader independently verifies every URL, quote, and claim against a detailed rubric. The grader provides structured feedback that drives revisions until the brief passes, eliminating manual review cycles. Key techniques include writing specific, actionable rubrics that force concrete evidence, using span.outcome_evaluation_* events to track the loop, and understanding when Outcomes is the right tool for quality assurance.

This cookbook demonstrates building a multiagent coordinator system using Claude Managed Agents to automate sales-proposal generation. A coordinator agent orchestrates three specialist subagents—a web-search researcher, a case-study librarian, and a rules-based pricing modeler—each with scoped toolsets and specific responsibilities. The pattern shows how to structure heterogeneous teams, use per-role tool scoping, and leverage thread_created/thread_message_received events for agent coordination to assemble tailored sales proposals.

OpenAI Agents Python SDK v0.16.0 introduces significant updates including a default model change from gpt-4.1 to gpt-5.4-mini, new configuration options for disabling max_turns limits, SDK-side tool execution concurrency control, and server-prefixed MCP tool naming to prevent conflicts. The release also includes stability fixes for tool call output indexes and security improvements for symlink handling. These changes enhance performance and provide greater flexibility for agent configuration while maintaining backward compatibility through explicit model specification.

LangChain v1.3.0a2 is an alpha release featuring improvements to stream events protocol (v3), ordered schema resolution, and agent middleware enhancements. The release includes performance optimizations, bug fixes across multiple integrations (OpenAI, Anthropic, Fireworks), and new features like HITL middleware respond decision and dynamic tool registration. Key updates involve dependency bumps, security patches, and refinements to the agent creation and model initialization systems.

crewAI v1.14.5a3 is a maintenance release focusing on bug fixes and architectural improvements. Key changes include correcting the status endpoint path structure, upgrading gitpython for security compliance, and extracting the CLI into a standalone package. This refactoring improves modularity while maintaining backward compatibility.

OpenAI Agents Python v0.15.2 introduces context management model settings and fixes critical issues with conversation session replay, tool execution, and error handling. The release includes 11 bug fixes addressing assistant conversation item IDs, function tool tracing, MCP tool metadata isolation, and stream terminal rejection. Comprehensive test coverage improvements ensure reliability across realtime tool behaviors, handoff mechanisms, and tool identity helpers.

This cookbook demonstrates building a vulnerability-discovery agent using the Claude Agent SDK that automatically threat-models C source code, hunts memory-safety bugs using built-in file tools (Read, Grep, Glob), and generates structured security reports. The agent operates in a multi-turn session with a bootstrap threat-modeling phase, an interview phase for owner input, and automated vulnerability finding and triage loops. The approach reduces false positives compared to traditional static analyzers by using Claude's reasoning to identify high-confidence memory-safety issues in a read-only sandbox environment.

crewAI v1.14.5a2 is a bug fix release addressing critical issues in task output handling, token counting, async operations, and LLM configuration. Key improvements include fixing task output restoration in finally blocks, preserving outputs across async batch flushes, and preventing incorrect final answer returns from hooks and errors. The release also enhances the CrewAIRagAdapter with proper kwargs forwarding and improves BaseModel input handling.

LangChain v1.3.0a1 is an alpha release featuring stream_events(version='v3') protocol integration into create_agent, HITL middleware enhancements with a 'respond' decision, and performance optimizations including reduced agent state inlining and LRU caching for type resolution. The release includes updates across multiple partner packages (Fireworks, Anthropic, OpenAI), dependency bumps for security (Pygments CVE-2026-4539, cryptography), and improvements to token counting, middleware initialization speed, and agent configuration. Key infrastructure changes include content-block-centric streaming (v2), dynamic tool registration via middleware, and enhanced LangSmith integration metadata.

crewAI v1.14.5a1 introduces the `restore_from_state_id` kickoff parameter for state management and enhances the ExaSearchTool with highlights while renaming it from EXASearchTool. Bug fixes address missing crewai pin sites in the release flow and ensure proper skills loading events for traces. This alpha release includes documentation updates and contributions from five team members.

OpenAI Agents Python v0.15.0 introduces explicit model refusal handling through a new `ModelRefusalError` exception, replacing previous behavior where refusals were treated as empty output or caused retry loops. The release includes a `model_refusal` error handler that allows developers to gracefully handle refusals without raising exceptions. For structured-output agents, handlers can return values matching the output schema for validation. Additional updates include documentation improvements and dependency bumps.

OpenClaw v2026.4.29-beta.4 introduces significant enhancements across messaging, memory, provider coverage, and reliability. Key additions include active-run steering by default, a people-aware wiki with provenance views, NVIDIA provider onboarding, and improved gateway/channel reliability. Security improvements include OpenGrep scanning and sharper GHSA triage policies, while multiple channel fixes address Slack, Telegram, Discord, WhatsApp, and Teams edge cases.

OpenClaw v2026.4.29 is a major release introducing active-run steering for messaging, people-aware memory wiki with provenance tracking, expanded provider coverage (NVIDIA, Bedrock), and enhanced reliability across gateway, channels, and security. Key improvements include opt-in follow-up commitments with heartbeat delivery, per-conversation Active Memory filters, faster model manifest paths, and comprehensive channel fixes for Slack, Telegram, Discord, WhatsApp, and Teams. Security enhancements add OpenGrep scanning and safer exec/pairing handling.

crewAI v1.14.4 introduces enhanced persistence customization, expanded cloud provider support (Azure OpenAI, Vertex AI), and new research tools (Tavily, You.com MCP). The release includes 12 bug fixes addressing JSON handling, message preservation, credential forwarding, and multimodal input support. Documentation updates cover E2B and Daytona sandbox tools.

OpenClaw v2026.4.29-beta.2 introduces significant enhancements across messaging, memory, provider coverage, and reliability. Key additions include active-run steering for messaging, a people-aware wiki with provenance tracking, NVIDIA provider onboarding, and improved channel resilience across Slack, Telegram, Discord, and others. Security improvements include OpenGrep scanning and stricter tool-scope handling, while operational reliability focuses on slow-host startup diagnostics and session recovery.

crewAI v1.14.4a1 is a pre-release update focusing on bug fixes and documentation improvements. Key fixes address LLM chat failures, message state management in executors, trained-agent file handling, multimodal input support, and JSON serialization issues. The release includes new documentation for sandbox tools (E2B, Daytona), Vertex AI setup, and You.com MCP integration, plus security updates to litellm.

OpenClaw v2026.4.26 introduces major enhancements to real-time communication, plugin architecture, and migration tooling. Key additions include a generic browser real-time transport contract, Cerebras provider integration, asymmetric embedding support for memory search, and bundled importers for Claude and Hermes configurations. The release also includes numerous stability fixes for gateway operations, plugin discovery, and device token handling.

This guide demonstrates how to build Claude Managed Agents with persistent memory capabilities that learn and retain user preferences across multiple sessions. Users create a memory store, configure a shopping agent with file access tools, and show how the agent automatically reads and updates customer preferences on repeat visits. The memory system uses a simple file-based approach mounted at /mnt/memory/ that both the agent and application can access via API.

OpenClaw v2026.4.24 introduces Google Meet as a bundled participant plugin with real-time voice capabilities, adds DeepSeek V4 models to the default catalog, and enhances browser automation with coordinate clicks and improved action budgets. The release significantly lightens plugin and model infrastructure at startup through static catalogs and lazy dependencies, while implementing breaking changes to plugin SDK tool-result transforms that require migration to the new `registerAgentToolResultMiddleware` API.

OpenClaw v2026.4.24-beta.1 introduces Google Meet as a bundled participant plugin with real-time voice capabilities, adds DeepSeek V4 models to the default catalog, and implements WebRTC-backed realtime voice loops across Talk, Voice Call, and Google Meet. The release includes browser automation improvements (coordinate clicks, longer action budgets, per-profile headless overrides), lighter plugin/model infrastructure at startup, and breaking changes to plugin SDK tool-result transforms requiring migration to `registerAgentToolResultMiddleware`.