Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 327 Agent status: 1 with issues Updated Oct 7, 2026
AgentGid

CrewAI changelog: what's new each month

Every stable CrewAI release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

September 2026

5 releases: 1.15.19 → 1.15.23

September brought improved platform integrations, better tracing and evaluation capabilities, and expanded model support. The month focused on reliability fixes and enhanced visibility into crew execution.

Highlights

  • Added support for Gemini 3.8 Flash and routed all DashScope models through the native provider.
  • Introduced evaluation of traced runs through the platform and added an Evaluate button to the TUI.
  • Enhanced tracing to capture task output formats, human feedback, and pause events for better debugging.
  • Improved platform integration setup with validation during crew setup and prioritized popular integrations.
  • Fixed multiple issues with URL reading, error reporting, and trace visibility.

New

  • Clipper integrations client for additional integration options.
  • Support for native Gemini 3.8 Flash model.
  • Telemetry tracking for checkpoint runtime and CLI usage.
  • Injectable client for CrewAI platform tools.
  • Platform tools support in JSON crew wizard.
  • CrewAI Platform application catalog exposure.
  • Support for aliases as connection identifiers.
  • llmoverlay context variable to route agent roles to models.

Improved

  • Enhanced tracing task spans to include declared output format and results.
  • Improved platform integration setup user experience with validation during crew setup.
  • Prioritized popular platform integrations in setup flow.
  • Now() function added to CEL expression environment.
  • Machine size reporting changed from core count to coarse bands.
  • Checkpoint persistence improved with UTF-8 encoding for JSON.
  • Human feedback and pause events now collected in tracing.
  • Task prompt and output carried in agent execution payloads.

Fixed

  • Fixed reading of octet-stream and xlsx URLs in urlreadtool.
  • Fixed appending of trailing user turn in native Gemini provider.
  • Fixed gateway errors incorrectly reported as HTTP 200 responses.
  • Fixed Oxylabs integration scrape failures that raised IndexError.
  • Fixed stray $ character in f-string search URLs for BrightData.
  • Fixed visibility of the panel for viewing traces and resolved issues with enabling tracing.

August 2026

8 releases: 1.15.11 → 1.15.18

August brought stability improvements and expanded capabilities for flows and conversational agents, with new tools for data access and enhanced telemetry tracking throughout the platform.

Highlights

  • Conversational flows promoted to stable with improved declaration support and documentation.
  • New URLReadTool added for reading arbitrary URLs, alongside IBM Db2 search tool.
  • Unified scaffolding under crewai create command for cleaner resource creation.
  • Enhanced telemetry now tracks flow outcomes, deployment origins, and execution context with UUID support.
  • Declarative conversational mode now drives flow behaviour with improved opt-in clarity.

New

  • URLReadTool for reading arbitrary URLs.
  • IBM Db2 search tool integration.
  • Execution context management with UUID support.
  • Declarative conversational flows that synthesize built-in methods.
  • Recording of exception types that end flows and trace batch sharing status.
  • Declaration ability to name router response formats and define chat flow state shape.
  • Project ID tracking for linking OSS usage to enterprise accounts.
  • FlowStartedEvent emission when boundary hooks abort a flow.

Improved

  • Unified scaffolding for all resources under crewai create command.
  • Conversational route and handler name collision errors now clarified.
  • LiteLLM-routed model provider preservation fixed.
  • Anthropic cache token usage now correctly reported.
  • CLI flags standardized to kebab-case format.
  • Oversized single messages during chunking now handled properly.
  • Failed turns no longer mark subsequent turns as failed.
  • Agent scope now closed on every failed attempt.

Fixed

  • LiteLLM provider preservation on routed models.
  • Underreporting of Anthropic cache token usage.
  • MySQL search table name validation.
  • URL hostname usage as MCP HTTP and SSE servername.
  • Broken README links, table of contents, and contribution guidance.
  • CEL error reporting for failures inside map literals.

July 2026

9 releases: 1.15.2 → 1.15.10

July brought major improvements to CrewAI's flow execution capabilities, skills repository, and tool handling. The month focused on making declarative flows production-ready while strengthening integration with the Skills Repository and improving reliability across LLM model support.

Action needed
  • E2BAPIKEY is now a required environment variable for E2B tools.

Highlights

  • Skills Repository is now out of experimental status and fully supported.
  • Declarative flows can now run on the terminal UI as a headless fallback.
  • New WaitTool added for managing long-running job pauses.
  • Execution interception points and hooks reworked for better control over agent behavior.
  • Tool calling improved across multiple LLM providers including GPT-5.6 and Anthropic models.

New

  • WaitTool for pausing execution during long-running jobs.
  • Flow Definition authoring skill for generated flows.
  • Execution step interception points with generic dispatcher.
  • Skill usage events emitted at runtime for observability.
  • Support for templated Flow action inputs.
  • Repository agents can be added to flow definitions.
  • Progressive disclosure UI for skills.

Improved

  • Latest LLM models now pulled dynamically in the crew wizard.
  • Inline skill definitions are now supported.
  • Execution hooks documentation reworked around the @on decorator.
  • Message setup and feedback handling improved in AgentExecutor.
  • Tool schema property names now preserved in strict mode.
  • Skill registry downloads now authenticated.
  • Model availability guidance refreshed.
  • Skill registry skills resolved through CrewAI+ client at runtime.

Fixed

  • Null repository agent attributes now handled correctly.
  • Anthropic preview tool-use blocks detection fixed.
  • Tool calling now works on the Responses API path.
  • Responses-only models no longer fail with 404 errors.
  • FileWriterTool write operations and rough edges addressed.
  • Tool failures now surfaced instead of reported as success.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. CrewAI: pricing, features and alternatives · All changelogs