deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

OpenAI Agents SDK closed 98% of new issues in 30 days; Anthropic's Claude Agent SDK 1 of 54

A dev.to study of 3,562 GitHub issues across nine AI agent frameworks found OpenAI's Agents SDK closed 98.3% within 30 days, while Anthropic's Claude Agent SDK closed 1 of 54.

OpenAI Agents SDK closed 98% of new issues in 30 days; Anthropic's Claude Agent SDK 1 of 54

A dev.to analysis published on November 21, 2026 set out to answer a practical question about AI agent frameworks: when you file a bug report, does anyone close it? Drawing on 3,562 human-filed GitHub issues, the study found a stark spread in responsiveness. OpenAI's Agents SDK for Python closed 98.3% of new issues within 30 days; Anthropic's Claude Agent SDK closed one issue out of 54.

How the measurement works

According to the dev.to post, the author built one cohort per repository: issues (not pull requests) opened between August 23 and October 22, 2026, with state read on November 21 through GitHub's search API. Bot-authored issues were removed, which took out 226 from Pydantic AI and 99 from Mastra, and the two busiest repositories had their windows split in halves to work around the API's 1,000-result cap. Microsoft's agent-framework repository was dropped after returning a 404 at collection time. Because every issue in the cohort had a full 30 days in which it could be closed, the 30-day figures are comparable across repositories; still-open issues count as never closed when computing medians.

The ranking

Repository Cohort issues Closed ≤ 30 days Median time to close
openai/openai-agents-python 181 98.3% 1.25 days
mastra-ai/mastra 1,396 86.6% 2.33 days
google/adk-python 291 66.0% 16.69 days
pydantic/pydantic-ai 811 52.4% 26.62 days
crewAIInc/crewAI 468 27.6% 36.73 days
langchain-ai/langgraph 207 17.4% not reached
microsoft/autogen (archived) 32 12.5% not reached
huggingface/smolagents 122 6.6% not reached
anthropics/claude-agent-sdk-python 54 1.9% not reached

Only two repositories closed even half of their new issues within a week: OpenAI's SDK and Mastra. The median OpenAI Agents SDK issue was closed in roughly 30 hours, and Mastra absorbed 7.7 times the issue volume while still closing its median issue in 2.33 days. At the other end, 40.7% of the Claude Agent SDK's 54 issues were still open with zero comments when the data was collected, with smolagents and the archived AutoGen repository at 24.6% and 34.4% respectively.

Stars and release cadence are poor proxies

Two of the most-starred active repositories performed worst on timely closures. CrewAI, with 61,134 stars, closed 27.6% of its cohort within 30 days; LangGraph, with 45,215 stars, closed 17.4%. The OpenAI SDK, carrying roughly half of CrewAI's star count, closed 98.3%.

Release frequency points in a similarly misleading direction. Citing a November 12 release-tracking read, the post notes that the Claude Agent SDK's latest stable tag was only three days old at the time, against a 1.9% 30-day closure rate, while the OpenAI SDK's stable tag was 28 days old alongside its 98.3% rate. Shipping often and triaging the public tracker are separate signals, the author argues.

Closed does not mean fixed

The study uses GitHub's closure-reason field to separate genuine triage from housekeeping. Of CrewAI's closed issues, 65.7% were marked not planned, as were 50.6% of LangGraph's; Mastra and Google ADK sat at 3.5% and 6.4%. Sampling eight closed issues per repository, the author found two of CrewAI's eight were shut by a stale-issue bot, and all eight sampled LangGraph closures were performed by a single account, five of them marked not planned. Every sampled OpenAI and Mastra closure came from a human account and was marked completed, though the author cautions that eight is too small a sample to read as a rate. CrewAI's eventual 70.9% closed-now figure also needs context: of its 332 closed issues, 203 were closed after the 30-day mark.

Why it matters

Framework selection usually weighs licence, features and release cadence, but this dataset argues for adding a maintenance-responsiveness column. If your team depends on upstream fixes, a vendor SDK that lets its tracker sit idle, or an open-source project whose closures are driven by stale bots, changes the real cost of adoption. The caveats are real: closed is not fixed, repositories that steer users toward forums or Discussions will look slower than they feel, and this covers a single 60-day window rather than a long-run trend. The author published the two-line GitHub query so the check can be rerun against whatever repository you actually pin.

  • #ai-agents
  • #open-source
  • #github
  • #openai
  • #anthropic

Related posts