Key findings
- Prototype speed earns the most consistent praise: 10 sources credit role-based crews with the fastest route to a working multi-agent workflow.
- Runtime defects lead the concerns at 10 sources, followed by coarse control and by defaults, telemetry, and dependencies at 9 each.
- Token accounting draws 7 sources, among them usage counts doubled for agents sharing one LLM instance and Bedrock counts that leave out cache reads.
- Tools that rerun on retry or never run at all appear in 7 sources.
- Of 12 sources that describe keeping or leaving CrewAI, 7 kept it and 5 moved away.
- Named organizations left in 2026: EpicStaff deleted its vendored CrewAI, Browserbase removed Stagehand's CrewAI integration as stale, and Headroom and ai-agent-assembly dropped it from their lockfiles over a critical ChromaDB advisory.
The public accounts behind CrewAI's Signals agree on what it does fastest and split on whether teams keep it. The English-language sources span January 2025 to October 2026, from Thoughtworks' Technology Radar, engineering posts by Diagrid and Vstorm, and pull requests in which EpicStaff, Browserbase and Headroom removed CrewAI, to GitHub issues and solo developers.
What builders keep it for
Builders name speed. A dev.to builder found it quick to prototype with byte-identical output across runs, and a builder on X compares it to managing a team. Vstorm's same-app comparison found CrewAI the longest version at about 420 lines, with research and writer roles ready out of the box.
Real work follows. Thoughtworks used CrewAI several times on client production problems, from promotion-code validation to transaction-failure investigation. A Hacker News commenter runs agents in production and contributed tools to the core repository, another ports agents into CrewAI on Amazon Bedrock AgentCore, and a Reddit builder runs an open-source stock research app on FastAPI and CrewAI.
CrewAI's own case studies add named companies with numbers but no method. PwC reports code-generation accuracy rising from 10% to 70%, Gelato carrier onboarding falling from 5 days to 10 minutes, and IBM Consulting two federal eligibility pilots run by Flows. GILD lists them as vendor customer stories and counts none of them.
Where it adds work
Control draws the widest concern. Agents left to collaborate drift from the spec, a delegated coworker ran an export tool the orchestrator never had, and several comparisons rate CrewAI lower on control than LangGraph. Diagrid's CTO finds its persistence short of durable execution.
GitHub issues add specific defects: async guardrail retries that crash, filed by an AB InBev engineer, memory that grows when a crew is created per request, and a failed task skipped after checkpoint restore.
Tokens, retries, and defaults
Cost reports center on counting. One issue shows usage_metrics reporting 310 tokens for a run that used 155, and another shows Bedrock leaving cache reads out of prompt tokens. A cost audit finds default context passing makes a five-agent crew send about six times the tokens of isolated tasks.
Defaults draw the rest: telemetry on by default and still sent after disabling in one report, a tracing prompt that added about 20 seconds per run, and a 699 MB install. During the March 2026 LiteLLM compromise, CrewAI pinned the last good version without a warning in the commit.
Why some teams moved on
Five sources moved away: to AutoGen, Spring AI, a C++ framework, or a small framework of their own, and one commenter dismissed CrewAI outright. Their reasons are limits, learning curve, and speed.
Teams outside the count left too. EpicStaff deleted its vendored CrewAI and mem0, 180,251 lines, in September 2026, and Browserbase removed Stagehand's CrewAI integration as stale in October. Headroom and ai-agent-assembly dropped CrewAI from their lockfiles because it requires ChromaDB 1.5.9, under a critical advisory with no patch. Pydantic's founder ruled CrewAI out over type safety and built Pydantic AI. Ask these questions before you choose.
- Does the work divide into roles, or is it a sequence your code should control?
- Which tools have side effects, and are they safe to run twice?
- Who checks token counts against the provider's bill?
- Above 50 executions a month, will you buy Enterprise or host it yourself?
Methodology
GILD synthesis of the 55 English-language public accounts in CrewAI's Signals, opened on October 28, 2026 and published between January 2025 and October 2026. They comprise 14 Hacker News comments, 20 GitHub issues on CrewAI's repository and 4 pull requests in other organizations' repositories, 3 Reddit posts and comments read from a stored copy because Reddit refused direct requests, 3 posts on X, 7 dev.to posts, 3 company publications, and 1 podcast episode. The 4 pull requests state no stance and count toward no figure, and the podcast is listed and not counted. CrewAI's community forum, blog, and case studies are left out, since they sit on the vendor's domain. Counts describe the sources reviewed and say nothing about CrewAI users at large. Product behavior is checked against CrewAI's documentation on the same date.
Products in this piece
Sources
- Public source
- 26
- Official documentation
- 8
- Thoughtworks Technology Radar: CrewAI, placed in Assess (Apr 2, 2025)
- Kacper Wlodarczyk (Vstorm): One Chat App, 4 AI Frameworks Compared (Mar 13, 2026)
- Yaron Schneider (Diagrid): Why Checkpoints Aren't Durable Execution, LangGraph, CrewAI, Google ADK, and Others (Feb 25, 2026)
- sunnydachs on dev.to: The same agent in Strands, LangGraph, and CrewAI with every LLM call recorded (Sep 21, 2026)
- wartzarbee on dev.to: CrewAI's quadratic context problem, why a 5-agent crew costs 6x more (Jul 26, 2026)
- Rafael Asor on dev.to: I delegated a review in CrewAI, and the reviewer ran an export tool the orchestrator never had (Sep 3, 2026)
- Hacker News comment on CrewAI agents collaborating awkwardly without a spec (Sep 9, 2026)
- Hacker News comment on running CrewAI agents in production and contributing tools (Dec 14, 2025)
- Hacker News comment on a local CrewAI, ChromaDB, and Ollama setup (Jan 8, 2025)
- kingabzpro on Reddit: an open-source stock research app on FastAPI and CrewAI (Oct 24, 2026)
- Hacker News comment on leaving CrewAI and LangGraph for a small framework (Sep 25, 2025)
- Hacker News comment on rebuilding a research tool from CrewAI on AutoGen (May 3, 2025)
- CrewAI issue #7259: usage_metrics multiplied by agents sharing an LLM instance (Sep 4, 2026)
- CrewAI issue #8626: Bedrock prompt_tokens excludes cache reads (Oct 26, 2026)
- CrewAI issue #7252: kickoff_async guardrail retry crashes, filed by an AB InBev engineer (Sep 3, 2026)
- CrewAI issue #7338: memoized CrewBase methods keep every crew in a global cache (Sep 8, 2026)
- CrewAI issue #8718: a failed task skipped after checkpoint restore (Oct 28, 2026)
- CrewAI issue #2945: telemetry still sent after turning it off (Jun 3, 2025)
- CrewAI issue #3789: tracing prompt adds about 20 seconds per run (Oct 24, 2025)
- Hacker News comment on CrewAI pinning LiteLLM 1.82.6 during the March 2026 compromise (Mar 24, 2026)
- EpicStaff pull request #768: delete vendored CrewAI and mem0 libraries (Aug 20, 2026, merged Sep 4, 2026)
- Browserbase Stagehand pull request #3336: remove the stale CrewAI integration (Oct 27, 2026)
- Headroom pull request #3870: drop the crewai extra to remove chromadb from the lockfile (Sep 30, 2026)
- ai-agent-assembly pull request #133: remove crewai to fix a critical chromadb vulnerability (Jul 7, 2026)
- GitHub advisory GHSA-f4j7-r4q5-qw2c: ChromaDB pre-authentication code injection, critical (May 18, 2026)
- Software Engineering Daily: Pydantic AI with Samuel Colvin (Dec 4, 2025)
- CrewAI customer story: PwC accelerates enterprise-scale GenAI adoption with CrewAI (undated, read Oct 28, 2026)
- CrewAI customer story: Gelato accelerates fulfillment via agentic integration (undated, read Oct 28, 2026)
- CrewAI customer story: IBM automates federal eligibility with agents (undated, read Oct 28, 2026)
- CrewAI docs: Introduction (Crews and Flows)
- CrewAI docs: Flows (state and @persist)
- CrewAI docs: Checkpointing
- CrewAI docs: Telemetry
- CrewAI pricing (AMP Basic and Enterprise, executions, overage)