CrewAI−7 Mixed

Operator Insights

What are teams learning from building on CrewAI and running it in production?

10 accounts: 2 failures, incidents or rollbacks, 3 left it or turned it down, 2 evaluated it, 2 adopted it and 1 reported by others

Failures, incidents and rollbacks2

Jucelio QuentinoDeveloperAB InBev
CrewAI agents run with kickoff_async and an output guardrail that can reject and retry.
GitHub (crewAIInc/crewAI), published Sep 3, 2026Read the issue (opens GitHub (crewAIInc/crewAI) in a new tab)
Observation
An AB InBev developer found that async CrewAI agents crashed on their first guardrail rejection and never retried.
Takeaway
Test guardrail retries on the async path you run in production, since a rejection there can fail the run before a retry.
What happened
When a guardrail rejected an async agent's first output, CrewAI sent the retry through the synchronous executor inside the event loop. The executor returned a coroutine where a result was expected, and the run crashed before the retry could complete, so guardrail_max_retries did not work for async agents on crewai 1.15.18. The issue is still open.
- Public engineering write-up (the engineer's GitHub issue from a corporate account)
- No team metrics reported
Deployment details
CrewAI agents run with kickoff_async and an output guardrail that can reject and retry.
- Public engineering write-up (the engineer's GitHub issue from a corporate account)
- No team metrics reported
Team
AB InBev
Workload
Async agents with output guardrails
Product version
crewai 1.15.18
In SignalsCounted once, as the engineer's own issue. In the record

ThoughtworksApr 15, 2026GitHub (crewAIInc/crewAI)A Thoughtworks developer found that since CrewAI 1.9.0 structured task outputs stopped tools from running on vLLM-served models.
Arvind ThangamaniDeveloperThoughtworks
CrewAI tasks with output_pydantic on models served through vLLM and other OpenAI-compatible servers.
GitHub (crewAIInc/crewAI), published Apr 15, 2026Read the issue (opens GitHub (crewAIInc/crewAI) in a new tab)
Observation
A Thoughtworks developer found that since CrewAI 1.9.0 structured task outputs stopped tools from running on vLLM-served models.
Takeaway
Retest tool-calling agents with structured outputs after every CrewAI upgrade, especially on self-served models.
What happened
From v1.9.0 CrewAI passed a task's output_pydantic into the agent's tool-calling loop as a response format on every step. Models served through vLLM then skipped tool calls and returned structured JSON directly, so the tools never ran. Before 1.9.0 the format was applied after the tools finished, and the report calls the change a regression. The issue is closed.
Public engineering write-up (the developer's GitHub issue)
No team metrics reported
Deployment details
CrewAI tasks with output_pydantic on models served through vLLM and other OpenAI-compatible servers.
Public engineering write-up (the developer's GitHub issue)
No team metrics reported
Team
Thoughtworks
Workload
Tool-calling agents with structured task outputs
Product version
crewai 1.9.0 and later
Model
Models served through vLLM
In SignalsNot counted. Not one of the coded Signals sources for this Hub.

Left it or turned it down3

Vitalii Povoroznyk (VitaliiPovoroznykHYS)Contributor (pull request author)EpicStaff
EpicStaff, a source-available, self-hosted platform for building agent flows in a visual editor over a Django backend, which had carried CrewAI inside its own code.
GitHub, EpicStaff/EpicStaff, published Aug 20, 2026Read the write-up (opens GitHub, EpicStaff/EpicStaff in a new tab)
Observation
EpicStaff dropped CrewAI from its self-hosted agent platform, deleting its vendored copies of CrewAI and mem0 in one change of 180,251 lines.
Takeaway
Vendoring a framework buys control at the cost of carrying it. Once a platform has its own agent model, budget for removing the framework, and keep CrewAI behind an interface you can replace.
What happened
Pull request #768, opened Aug 20, 2026 and merged Sep 4, 2026, removed the CrewAI library as an EpicStaff dependency. It deleted the vendored CrewAI and mem0 libraries, 180,251 lines across 1,403 files, and replaced the old project and crew screens with agent definitions. EpicStaff builds its own agent-flow platform, an alternative to CrewAI's.
- Public engineering write-up (GitHub pull request)
- Team metrics reported
Deployment details
EpicStaff, a source-available, self-hosted platform for building agent flows in a visual editor over a Django backend, which had carried CrewAI inside its own code.
- Public engineering write-up (GitHub pull request)
- Team metrics reported
Team
EpicStaff platform engineers
Workload
Agent flows built in EpicStaff's visual editor
Product version
A vendored copy of CrewAI inside EpicStaff
Measurement
180,251 lines deleted across 1,403 files
In SignalsNot counted. EpicStaff builds a competing agent platform, so its pull request is listed in Signals and not counted.

PydanticDec 4, 2025Software Engineering DailyPydantic's founder looked at CrewAI, LangGraph and other frameworks for Logfire's AI features, ruled them out over type safety, and built Pydantic AI.
Samuel ColvinFounderPydantic
Pydantic's Logfire observability product, choosing a framework for its first AI features in late 2024.
Software Engineering Daily, published Dec 4, 2025Listen to the episode (opens Software Engineering Daily in a new tab)
Observation
Pydantic's founder looked at CrewAI, LangGraph and other frameworks for Logfire's AI features, ruled them out over type safety, and built Pydantic AI.
Takeaway
If type safety and library engineering standards gate what you adopt, check them in CrewAI's code before the first crew, because a competitor's founder ruled out CrewAI on exactly those grounds.
What happened
On Software Engineering Daily (episode of Dec 4, 2025), Colvin stated that when Pydantic began building AI functionality into Logfire he expected one of the frameworks built on Pydantic, among them LangChain, LangGraph, CrewAI and LlamaIndex, to fit. He was disappointed: they were not type safe and fell short of the engineering standard of the top Python packages, so the team built Pydantic AI. Pydantic AI now competes with CrewAI.
“Obviously, LangChain, LangGraph, Crew AI, LlamaIndex, all of these guys used Pydantic. And I assumed there was going to be a good one that I could go and use. Started looking at them and was super disappointed by what I found.”
Samuel Colvin, at 13:08 in the recording
Podcast episode (Software Engineering Daily, publisher transcript)
No team metrics reported
Deployment details
Pydantic's Logfire observability product, choosing a framework for its first AI features in late 2024.
Podcast episode (Software Engineering Daily, publisher transcript)
No team metrics reported
Team
Pydantic's Logfire team
Workload
AI features for the Logfire observability product
In SignalsNot counted. Its speaker founded Pydantic, whose Pydantic AI competes with CrewAI, so the record is listed and not counted.
mahdiyarSep 25, 2025Hacker NewsThe builder of AskBudi's tinyAgent tried CrewAI and LangGraph, found the learning curve more than the features he used needed, and chose to write his own framework instead.
mahdiyarBuilder of tinyAgent and RoundtableAskBudi
Trying agent frameworks for AskBudi's own tools, from about September 2024.
Hacker News, published Sep 25, 2025Read the write-up (opens Hacker News in a new tab)
Observation
The builder of AskBudi's tinyAgent tried CrewAI and LangGraph, found the learning curve more than the features he used needed, and chose to write his own framework instead.
Takeaway
Count the CrewAI features your crew uses before the framework becomes a dependency to learn. Small agent loops are short enough to own.
What happened
On Hacker News on Sep 25, 2025, mahdiyar wrote that he started a year earlier by playing with CrewAI, LangGraph and other frameworks, found the learning curve more than necessary for the features he used, and wrote tinyAgent, a small framework of his own.
Public practitioner write-up (builder's Hacker News comment)
No team metrics reported
Deployment details
Trying agent frameworks for AskBudi's own tools, from about September 2024.
Public practitioner write-up (builder's Hacker News comment)
No team metrics reported
Team
AskBudi's builder
Workload
Exploratory agent tools
In SignalsCounted once, as the builder's own comment. In the record

Evaluated it2

Kacper WłodarczykAI EngineerVstorm
One full-stack chat app with research and writer roles, rebuilt in Pydantic AI, LangChain, LangGraph, and CrewAI.
Vstorm open source blog, published Mar 13, 2026Read the write-up (opens Vstorm open source blog in a new tab)
Observation
Vstorm built the same chat app in four frameworks and evaluated CrewAI against Pydantic AI, LangChain, and LangGraph.
Takeaway
Count the glue code streaming and async need before choosing CrewAI for an app that serves users in real time.
What happened
The CrewAI version came with research and writer roles out of the box and took about 420 lines, against about 160 for Pydantic AI. It needed a background thread for streaming and run_in_executor for async calls.
- Public engineering write-up (the engineer's comparison on Vstorm's blog)
- Team metric reported
Deployment details
One full-stack chat app with research and writer roles, rebuilt in Pydantic AI, LangChain, LangGraph, and CrewAI.
- Public engineering write-up (the engineer's comparison on Vstorm's blog)
- Team metric reported
Team
Vstorm open-source team
Workload
A full-stack chat app built four times
Measurement
About 420 lines for CrewAI against about 160 for Pydantic AI
In SignalsCounted once, as the engineer's own post. In the record

IBMOct 28, 2026CrewAI customer storyAn IBM Consulting team ran two CrewAI pilots inside US federal agencies for eligibility decisions, with Flows controlling when rules or agents run.
Narayanan AlavandarQuoted in CrewAI's case studyIBM
IBM Consulting's automation work for US federal agencies, moving eligibility processing off legacy RPA, on IBM's watsonx models.
CrewAI customer story, published Oct 28, 2026Read the write-up (opens CrewAI customer story in a new tab)
Observation
An IBM Consulting team ran two CrewAI pilots inside US federal agencies for eligibility decisions, with Flows controlling when rules or agents run.
Takeaway
Where rules decide the outcome, let Flows hold the rules and give agents only the extraction and summary steps.
What happened
CrewAI's undated case study, read Oct 28, 2026, says the team began evaluating multi-agent frameworks in 2024 and by 2025 had two CrewAI pilots running inside federal agencies. Agents extract applicant data from documents, summarize findings and calculate eligibility, inside a hybrid of Flows and crews that decides when rules or agents run. Early pilots were faster than legacy RPA, with no figures given, and IBM Consulting was finalizing enterprise licenses to move them into production.
“By 2025 his team had two CrewAI pilots running inside federal agencies…”
CrewAI case study, on IBM Consulting
Vendor customer story (CrewAI case study, undated, read Oct 28, 2026)
No team metrics reported
Deployment details
IBM Consulting's automation work for US federal agencies, moving eligibility processing off legacy RPA, on IBM's watsonx models.
Vendor customer story (CrewAI case study, undated, read Oct 28, 2026)
No team metrics reported
Team
IBM Consulting federal automation team
Duration
Evaluation from 2024; two pilots by 2025
Workload
Applicant eligibility decisions for federal agencies
Product version
CrewAI Flows and crews on watsonx
In SignalsNot counted. A vendor case study on CrewAI's own site, so GILD does not count it as a Signals source.

Adopted it2

PwCOct 28, 2026CrewAI customer storyPwC uses CrewAI agents to generate and validate code in a proprietary language and to draft long specifications, and reports code-generation accuracy rising from 10% to 70%.
Jacob WilsonQuoted in CrewAI's case studyPwC
PwC's enterprise GenAI work, re-engineering software delivery workflows for consultants.
CrewAI customer story, published Oct 28, 2026Read the write-up (opens CrewAI customer story in a new tab)
Observation
PwC uses CrewAI agents to generate and validate code in a proprietary language and to draft long specifications, and reports code-generation accuracy rising from 10% to 70%.
Takeaway
Ask for the evaluation set behind a vendor's accuracy figure, and track agent choices and run times from day one so your own numbers can stand up.
What happened
CrewAI's undated case study, read Oct 28, 2026, says PwC re-engineered its software delivery workflows with CrewAI agents that generate, run and validate proprietary-language code, and embedded agents to draft functional and technical specifications with consultant feedback. CrewAI's monitoring tracked agent choices, run times and accuracy for leadership dashboards, and the case study reports code-generation accuracy rising from 10% to 70%. It does not say how the accuracy was measured.
“Crew-powered agents boosted code-generation accuracy from 10% to 70%, slashed turnaround time on complex documents, and supplied granular data to prove ROI.”
CrewAI case study, on PwC
Vendor customer story (CrewAI case study, undated, read Oct 28, 2026)
Team metrics reported
Deployment details
PwC's enterprise GenAI work, re-engineering software delivery workflows for consultants.
Vendor customer story (CrewAI case study, undated, read Oct 28, 2026)
Team metrics reported
Team
PwC consultants and engineers
Workload
Proprietary-language code generation and specification drafting
Product version
CrewAI agents with CrewAI's monitoring
In SignalsNot counted. A vendor case study on CrewAI's own site, so GILD does not count it as a Signals source.
GelatoOct 28, 2026CrewAI customer storyGelato runs CrewAI agents behind Gelato Connect that write, test and deploy carrier-integration code, and reports carrier onboarding falling from 5 days to 10 minutes.
Styrbjörn HolmbergQuoted in CrewAI's case studyGelato
Gelato, a print-on-demand platform with 250 staff across 50 countries, scaling without adding headcount in step.
CrewAI customer story, published Oct 28, 2026Read the write-up (opens CrewAI customer story in a new tab)
Observation
Gelato runs CrewAI agents behind Gelato Connect that write, test and deploy carrier-integration code, and reports carrier onboarding falling from 5 days to 10 minutes.
Takeaway
Agents that write and deploy integration code need tests in the loop and a person who owns each deploy. Ask how Gelato reviews what its agents ship before you copy the pattern.
What happened
CrewAI's undated case study, read Oct 28, 2026, says Gelato embedded thousands of CrewAI agents behind Gelato Connect to map bulk product catalogs and validate file uploads, and deployed logistics agents that generate, test and deploy carrier-integration code. It reports carrier onboarding falling from 5 days to 10 minutes and SKU-mapping time falling by more than 90%.
“Deployed logistics agents that generate, test, and deploy carrier‑integration code, shrinking onboarding from 5 days to 10 minutes.”
CrewAI case study, on Gelato
Vendor customer story (CrewAI case study, undated, read Oct 28, 2026)
Team metrics reported
Deployment details
Gelato, a print-on-demand platform with 250 staff across 50 countries, scaling without adding headcount in step.
Vendor customer story (CrewAI case study, undated, read Oct 28, 2026)
Team metrics reported
Team
Gelato product and logistics engineering
Workload
Catalog mapping, file validation and carrier-integration code
Product version
CrewAI agents and internal crews
Measurement
Carrier onboarding from 5 days to 10 minutes; SKU mapping more than 90% faster
In SignalsNot counted. A vendor case study on CrewAI's own site, so GILD does not count it as a Signals source.

Reported by others1

DiagridFeb 25, 2026Diagrid blogDiagrid's CTO evaluated CrewAI's persistence and found it stops short of durable execution for production workflows.
Yaron SchneiderCTO and Co-FounderDiagrid
Production agent workflows that must survive crashes and resume where they stopped.
Diagrid blog, published Feb 25, 2026Read the write-up (opens Diagrid blog in a new tab)
Observation
Diagrid's CTO evaluated CrewAI's persistence and found it stops short of durable execution for production workflows.
Takeaway
Decide who resumes a failed CrewAI flow before production, since persistence saves state but does not restart work.
What happened
Diagrid, which sells a durable workflow runtime, evaluated how agent frameworks recover from failures. For CrewAI it found that task replay keeps only the last kickoff and that @persist saves flow state to SQLite without resuming on its own, so the team writes recovery itself.
Public organization write-up (CTO's post, vendor of a competing runtime)
No team metrics reported
Deployment details
Production agent workflows that must survive crashes and resume where they stopped.
Public organization write-up (CTO's post, vendor of a competing runtime)
No team metrics reported
Team
Diagrid
Workload
Evaluation of recovery in agent frameworks
In SignalsCounted once, as the CTO's own post. In the record

Failures, incidents and rollbacks2

An AB InBev developer found that async CrewAI agents crashed on their first guardrail rejection and never retried.

A Thoughtworks developer found that since CrewAI 1.9.0 structured task outputs stopped tools from running on vLLM-served models.

Left it or turned it down3

EpicStaff dropped CrewAI from its self-hosted agent platform, deleting its vendored copies of CrewAI and mem0 in one change of 180,251 lines.

Pydantic's founder looked at CrewAI, LangGraph and other frameworks for Logfire's AI features, ruled them out over type safety, and built Pydantic AI.

The builder of AskBudi's tinyAgent tried CrewAI and LangGraph, found the learning curve more than the features he used needed, and chose to write his own framework instead.

Evaluated it2

Vstorm built the same chat app in four frameworks and evaluated CrewAI against Pydantic AI, LangChain, and LangGraph.

An IBM Consulting team ran two CrewAI pilots inside US federal agencies for eligibility decisions, with Flows controlling when rules or agents run.

Adopted it2

PwC uses CrewAI agents to generate and validate code in a proprietary language and to draft long specifications, and reports code-generation accuracy rising from 10% to 70%.

Gelato runs CrewAI agents behind Gelato Connect that write, test and deploy carrier-integration code, and reports carrier onboarding falling from 5 days to 10 minutes.

Reported by others1

Diagrid's CTO evaluated CrewAI's persistence and found it stops short of durable execution for production workflows.