Operator Insights

What are teams learning from building on OpenAI Agents SDK and running it in production?

10 accounts: 1 failure, incident or rollback, 1 left it or turned it down, 1 switched, 6 adopted it and 1 reported by others

Failures, incidents and rollbacks1

Loïc MinaudierSoftware developerDatadog
Agents on GPT-6 Sol with function and MCP tools, called through Chat Completions behind an OpenAI-compatible gateway.
GitHub (openai/openai-agents-python), published Oct 8, 2026Read the issue (opens GitHub (openai/openai-agents-python) in a new tab)
Observation
A Datadog developer's agents on GPT-6 Sol failed on every function tool call until the SDK learned GPT-6 reasoning defaults.
Takeaway
Test each new model family with tools before switching the agent's model, since the SDK's defaults are tuned per family.
What happened
The SDK applied default reasoning settings only to GPT-5 model names, so GPT-6 agents were sent no reasoning setting. On Chat Completions, GPT-6 Sol then rejected every function tool call with a 400 error, while the same agent code worked on gpt-5.6-luna. The report traced the cause to the default model settings in openai-agents 0.19.2, and the issue is now closed.
- Public engineering write-up (the developer's GitHub issue)
- No team metrics reported
Deployment details
Agents on GPT-6 Sol with function and MCP tools, called through Chat Completions behind an OpenAI-compatible gateway.
- Public engineering write-up (the developer's GitHub issue)
- No team metrics reported
Team
Datadog, through an OpenAI-compatible gateway
Workload
Agents with function and MCP tools
Product version
openai-agents 0.19.2 on Python 3.12
Model
GPT-6 Sol
In SignalsCounted once, as the developer's own issue. In the record

Left it or turned it down1

jamesbriggsMaintainer of graphaiAurelio AI
Aurelio Labs' agent applications, which needed async code throughout and streaming.
Hacker News, published Oct 8, 2025Read the write-up (opens Hacker News in a new tab)
Observation
Aurelio Labs used the OpenAI Agents SDK, LangChain and LangGraph extensively, then left all three for graphai, an async framework of its own.
Takeaway
A small runtime is still a runtime to work around. When async and streaming carry the design, a thin loop your team owns can be cleaner than any framework.
What happened
On Hacker News on Oct 8, 2025, jamesbriggs wrote that the team had used LangChain, LangGraph and the Agents SDK pretty extensively and now builds on graphai, an internal framework that borrows LangGraph's graph approach, runs everything async and handles streaming. He calls the switch a much better and cleaner experience, and says graphai fits Aurelio's own use cases more than anyone else's.
- Public practitioner write-up (maintainer's Hacker News comment)
- No team metrics reported
Deployment details
Aurelio Labs' agent applications, which needed async code throughout and streaming.
- Public practitioner write-up (maintainer's Hacker News comment)
- No team metrics reported
Team
Aurelio Labs
Workload
Agent applications with async execution and streaming
In SignalsNot counted. Its author maintains graphai, a competing agent framework, so its record is listed and counts toward no figure.

Switched to or from it1

Sibidharan (sibidharan)EdTech builder; founder of Selfmade Ninja Academy by his own GitHub profile (self-described)Selfmade Ninja Academy
EdTech apps with AI learning and evaluation tools, where agents run as separate Python processes that stream to the front end over SSE or WebSockets.
Hacker News, published May 29, 2026Read the write-up (opens Hacker News in a new tab)
Observation
An EdTech builder migrated every one of his older Completions and Responses API projects to the OpenAI Agents SDK and finds it stable.
Takeaway
Teams leaving OpenAI's older APIs can migrate app by app: the SDK keeps sessions and token tracking in your own process and runs other providers' models.
What happened
On Hacker News on May 29, 2026, he wrote that after OpenAI pulled the plug on the Assistants API he moved all his old Completions and Responses API projects to the SDK, and that they feel good and stable. He relies on it for sessions, token and cache tracking, and streaming tool calls, and runs it against other inference providers and local models in LM Studio.
- Public practitioner write-up (builder's Hacker News comment)
- No team metrics reported
Deployment details
EdTech apps with AI learning and evaluation tools, where agents run as separate Python processes that stream to the front end over SSE or WebSockets.
- Public practitioner write-up (builder's Hacker News comment)
- No team metrics reported
Team
One builder's own EdTech projects
Workload
Chat and evaluation tools for students, streamed over SSE and WebSockets
Model
OpenAI models, other inference providers and local models in LM Studio
In SignalsCounted once, as the builder's own comment. In the record

Adopted it6

HarveyDec 10, 2025Harvey blogHarvey built the evaluation system for its agentic legal search on the OpenAI Agents SDK's OpenTelemetry traces, scored offline in LangSmith.
Lysia Li, Varun Nair, Aaron Stern and co-authorsHarvey engineeringHarvey
Harvey's agentic search, in which an agent picks the legal knowledge sources to query and runs the retrieval calls behind a lawyer's research question.
Harvey blog, published Dec 10, 2025Read the write-up (opens Harvey blog in a new tab)
Observation
Harvey built the evaluation system for its agentic legal search on the OpenAI Agents SDK's OpenTelemetry traces, scored offline in LangSmith.
Takeaway
Treat the SDK's traces as the input to your evals: export them over OpenTelemetry to the evaluation tool you already run, and score tool choice as well as the answer.
What happened
Harvey's engineers built centralized evaluation infrastructure on the Agents SDK's OpenTelemetry traces and run offline evaluations in LangSmith, tracking hallucination, tool recall, retrieval recall, formatting and answer quality (Harvey blog, Dec 10, 2025). Its legal team tests each change with new query patterns. With that eval data, tool selection precision rose from near zero to 0.8 to 0.9, and complex queries that had stopped after one tool call now run 3 to 10 retrieval operations.
“Centralized evaluation infrastructure that we built using the OpenAI Agent SDK's OpenTelemetry traces enables us to run offline evaluations in LangSmith. We track metrics like hallucination, tool recall, retrieval recall, formatting, and answer quality.”
Harvey engineering
Public engineering write-up
Team metrics reported
Deployment details
Harvey's agentic search, in which an agent picks the legal knowledge sources to query and runs the retrieval calls behind a lawyer's research question.
Public engineering write-up
Team metrics reported
Team
Harvey search engineering, with Harvey's legal team testing
Workload
Agentic search over legal knowledge sources
Product version
OpenAI Agents SDK traces exported over OpenTelemetry to LangSmith
Measurement
Tool selection precision from near zero to 0.8 to 0.9; 3 to 10 retrieval operations on complex queries
In SignalsRecorded once in Signals as Harvey's own post. It takes no stance, so it moves no score. In the record
CoinbaseMar 11, 2025GitHub, coinbase/agentkitCoinbase Developer Platform added an OpenAI Agents SDK extension to AgentKit on Mar 11, 2025, so agents built on the SDK can hold crypto wallets.
Carson Roscoe (CarsonRoscoe)AgentKit contributor (pull request author)Coinbase
AgentKit, Coinbase Developer Platform's open-source toolkit for giving AI agents a crypto wallet and onchain actions.
GitHub, coinbase/agentkit, published Mar 11, 2025Read the write-up (opens GitHub, coinbase/agentkit in a new tab)
Observation
Coinbase Developer Platform added an OpenAI Agents SDK extension to AgentKit on Mar 11, 2025, so agents built on the SDK can hold crypto wallets.
Takeaway
Ship your tools as an SDK extension with example agents, and teams already on the SDK can adopt them without writing the glue.
What happened
Pull request #550, merged Mar 11, 2025, added a Python framework extension for the Agents SDK to AgentKit, tested with a chatbot that reported its wallet address and requested testnet funds from a faucet.
Public engineering write-up (GitHub pull request)
No team metrics reported
Deployment details
AgentKit, Coinbase Developer Platform's open-source toolkit for giving AI agents a crypto wallet and onchain actions.
Public engineering write-up (GitHub pull request)
No team metrics reported
Team
Coinbase Developer Platform AgentKit maintainers
Duration
From Mar 11, 2025
Workload
Wallet and onchain tools for agents, with an example chatbot
Product version
AgentKit's OpenAI Agents SDK extension for Python
In SignalsRecorded once in Signals as the pull request. It takes no stance, so it moves no score. In the record
FramebridgeJan 7, 2026GitHub (openai/openai-agents-python)Framebridge's e-commerce team runs agents that end on chosen tool calls, and works around the SDK to add one final answer.
Luke Bidigare-CurtisPrincipal Engineer and Team LeadFramebridge
Agents for Framebridge's e-commerce team that stop when specific tools are called.
GitHub (openai/openai-agents-python), published Jan 7, 2026Read the issue (opens GitHub (openai/openai-agents-python) in a new tab)
Observation
Framebridge's e-commerce team runs agents that end on chosen tool calls, and works around the SDK to add one final answer.
Takeaway
Decide early how an agent ends its run, since the SDK's stop rules decide whether the user gets a final answer after the last tool call.
What happened
The team was using stopAtToolNames in the agent's tool-use behavior to end a run when certain tools are called. It wanted one more generation after the final tool call with no further tool calls allowed. Until the SDK offered that, the team defined the last generation in the tool call's parameters, which made its tools harder to keep reusable across agents.
Public engineering write-up (the team lead's GitHub feature request)
No team metrics reported
Deployment details
Agents for Framebridge's e-commerce team that stop when specific tools are called.
Public engineering write-up (the team lead's GitHub feature request)
No team metrics reported
Team
Framebridge e-commerce team
Workload
Agents that end their run on specific tool calls
In SignalsNot counted. A feature request about one setting, so GILD does not count it as a Signals source.
Vijit and JackMar 18, 2026Hacker NewsLark's builders open-sourced Runtime Use, a sandbox runtime whose default agent runs on the OpenAI Agents SDK with shell and web search tools.
Vijit and JackBuilders of Runtime Use (Show HN authors)Lark
Lark, where agents run in sandboxes to generate end-to-end tests.
Hacker News, published Mar 18, 2026Read the write-up (opens Hacker News in a new tab)
Observation
Lark's builders open-sourced Runtime Use, a sandbox runtime whose default agent runs on the OpenAI Agents SDK with shell and web search tools.
Takeaway
The SDK supplies the agent loop and leaves the sandbox server, streaming and cancellation to you. Budget for that layer, or start from an open runtime that already wraps the SDK.
What happened
In a Show HN post on Mar 18, 2026, the team said getting the sandbox infrastructure right for Lark took significant effort, so it extracted that layer and open-sourced it as Runtime Use. The runtime starts an agent server inside a sandbox, streams outputs and files over a websocket, cancels long tasks, and by default runs an Agents SDK agent with shell and web search tools; one flag switches it to a Claude Code agent.
“By default it runs an OpenAI Agents SDK agent with shell + web search tools.”
Vijit and Jack
Public practitioner write-up (builders' Show HN post)
No team metrics reported
Deployment details
Lark, where agents run in sandboxes to generate end-to-end tests.
Public practitioner write-up (builders' Show HN post)
No team metrics reported
Team
Lark's builders, Vijit and Jack
Workload
Agents that generate end-to-end tests inside sandboxes
Product version
OpenAI Agents SDK as Runtime Use's default agent
In SignalsCounted once, as the builders' own post. In the record
ag_21Apr 26, 2025Hacker NewsOctagon AI open-sourced an MCP server of simulated venture investors that runs on the OpenAI Agents SDK.
ag_21Show HN authorOctagon AI
Octagon VC Agents, an open-source MCP server that simulates venture capital investors.
Hacker News, published Apr 26, 2025Read the write-up (opens Hacker News in a new tab)
Observation
Octagon AI open-sourced an MCP server of simulated venture investors that runs on the OpenAI Agents SDK.
Takeaway
The SDK's handoffs and tools are enough to ship a multi-agent product quickly. Plan the hosting and trace export around it before users depend on it.
What happened
In a Show HN post on Apr 26, 2025, Octagon AI said the server runs on the Agents SDK, which gave it multi-agent reasoning, context handoff between agents and tool use out of the box.
“It’s powered by OpenAI’s new Agents SDK, so you get true agentic workflows out of the box:…”
ag_21
Public practitioner write-up (Show HN post)
No team metrics reported
Deployment details
Octagon VC Agents, an open-source MCP server that simulates venture capital investors.
Public practitioner write-up (Show HN post)
No team metrics reported
Team
Octagon AI
Workload
Multi-agent VC simulations served over MCP
In SignalsCounted once, as the team's own post. In the record
Ramnique SinghApr 22, 2025Hacker NewsRowboat Labs runs its agents runtime and copilot as Python apps built on the OpenAI Agents SDK.
Ramnique SinghCo-founderRowBoat Labs
Rowboat, an open-source IDE for building and managing multi-agent systems, with a Next.js dashboard.
Hacker News, published Apr 22, 2025Read the comment (opens Hacker News in a new tab)
Observation
Rowboat Labs runs its agents runtime and copilot as Python apps built on the OpenAI Agents SDK.
Takeaway
Expect the SDK to pull the agent runtime into its own Python service beside the rest of the product, and plan packaging around that split.
What happened
Rowboat's dashboard is a Next.js app, while its agents runtime and its copilot are Python apps because they use the OpenAI Agents SDK. The team was working out how to bundle the pieces into one native app.
Public practitioner write-up (co-founder's Hacker News comment)
No team metrics reported
Deployment details
Rowboat, an open-source IDE for building and managing multi-agent systems, with a Next.js dashboard.
Public practitioner write-up (co-founder's Hacker News comment)
No team metrics reported
Team
RowBoat Labs
Workload
Agents runtime and copilot for a multi-agent IDE
Product version
OpenAI Agents SDK in Python apps
In SignalsCounted once, as the co-founder's own comment. In the record

Reported by others1

SuperserveMay 19, 2026Superserve blogSuperserve evaluated the sandbox clients the Agents SDK supports and found the tradeoffs missing from OpenAI's docs.
Pavitra BhallaFounderSuperserve
Choosing an execution backend for SDK sandbox agents among the hosted and local clients OpenAI ships.
Superserve blog, published May 19, 2026Read the write-up (opens Superserve blog in a new tab)
Observation
Superserve evaluated the sandbox clients the Agents SDK supports and found the tradeoffs missing from OpenAI's docs.
Takeaway
Treat the SDK's supported sandbox list as a starting point and test isolation, persistence, and cold starts on your own tasks before choosing one.
What happened
OpenAI ships sandbox clients for seven hosted options plus two local ones without benchmarking or ranking them. Superserve, which builds persistent sandboxes for agents and says it has a stake in the result, evaluated the options on isolation, persistence, pricing, GPU support, data residency, and volume mounts. It found that cold starts can dominate short tasks and that pricing is published unevenly.
Public organization write-up (founder's post, vendor of a competing sandbox)
No team metrics reported
Deployment details
Choosing an execution backend for SDK sandbox agents among the hosted and local clients OpenAI ships.
Public organization write-up (founder's post, vendor of a competing sandbox)
No team metrics reported
Team
Superserve
Workload
Evaluation of the sandbox clients the SDK supports
Product version
OpenAI Agents SDK sandbox clients
In SignalsCounted once, as Superserve's own post. In the record

Failures, incidents and rollbacks1

A Datadog developer's agents on GPT-6 Sol failed on every function tool call until the SDK learned GPT-6 reasoning defaults.

Left it or turned it down1

Aurelio Labs used the OpenAI Agents SDK, LangChain and LangGraph extensively, then left all three for graphai, an async framework of its own.

Switched to or from it1

An EdTech builder migrated every one of his older Completions and Responses API projects to the OpenAI Agents SDK and finds it stable.

Adopted it6

Harvey built the evaluation system for its agentic legal search on the OpenAI Agents SDK's OpenTelemetry traces, scored offline in LangSmith.

Coinbase Developer Platform added an OpenAI Agents SDK extension to AgentKit on Mar 11, 2025, so agents built on the SDK can hold crypto wallets.

Framebridge's e-commerce team runs agents that end on chosen tool calls, and works around the SDK to add one final answer.

Lark's builders open-sourced Runtime Use, a sandbox runtime whose default agent runs on the OpenAI Agents SDK with shell and web search tools.

Octagon AI open-sourced an MCP server of simulated venture investors that runs on the OpenAI Agents SDK.

Rowboat Labs runs its agents runtime and copilot as Python apps built on the OpenAI Agents SDK.

Reported by others1

Superserve evaluated the sandbox clients the Agents SDK supports and found the tradeoffs missing from OpenAI's docs.