Intel / repo

crewAI 1.15.34

Source: github.com
Author: joaomdmoura
Date: Published Nov 1, 2026

Context

Much of crewAI 1.15.34 is about making failures visible. crewAI is a Python framework for building AI agents that use tools and run together as crews and flows, and this release turns a long list of failure cases into named errors, warnings or explicit exit codes. A missing or placeholder API key, including for OpenAI-compatible providers, now raises MissingAPIKeyError. An empty answer from the model is retried once and then raises a typed error. crewai+ commands that need a login exit with code 1, and login now waits for the full lifetime of its device code before exiting with code 1 on timeout.

Two fixes change what happens to work already in progress. When an agent keeps repeating the same tool call, crewAI now warns and keeps the work done so far when it stops the run. When a model returns JSON that does not match the expected output model, crewAI now refactors it instead of failing, and a tool-using agent is asked for when it finishes. The release also honors a configured _window_size and warns when it falls back to a default window.

The one entry under breaking changes adds documentation: it records the removal of the 1.15.29 agent executor and adds migration notes. Separately, the release restores import paths that had been removed, now with deprecation warnings, so code that relied on those paths should load again while being told it needs updating.

Key takeaways · AI-distilled

crewAI lists one breaking change in 1.15.34: documentation, with migration notes, for the agent executor removed in 1.15.29.
A missing or placeholder API key now raises MissingAPIKeyError, and the same typed error now applies to OpenAI-compatible providers.
Agents now honor context_window_size and warn when falling back to a default window, and JSON that does not match the output model is refactored instead of failing.
LiteLLM tool-call and structured-output calls now report token usage, and LiteLLM errors keep the provider's original exception.
crewai update now prints upgrade commands instead of rewriting UV projects, and crewai create no longer offers Gemini models that have been shut down.

Terms in this piece · Glossary

structured output — Forcing a model's response to match a schema, so downstream code can parse it instead of guessing at prose.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.

Read the source github.com