AI SDK+36 Positive

Operator Insights

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

7 accounts: 2 failures, incidents or rollbacks, 2 switched and 3 adopted it

Failures, incidents and rollbacks2

Jason Safaiyeh (safaiyeh)Engineer running the AI SDK in a production APIChappie
A production API on AWS Fargate with a 2 GB memory limit that sends PDFs to models as AI SDK file parts. The issue names no company. Chappie is the company on the reporter's GitHub profile.
GitHub, vercel/ai, published Sep 1, 2026Read the write-up (opens GitHub, vercel/ai in a new tab)
Observation
In a production incident on an API Jason Safaiyeh runs, one 14 MB PDF sent through the AI SDK killed a 2 GB container in about 40 seconds and took about 100 requests with it.
Takeaway
Load-test the largest file a user can upload through the exact SDK version you ship, and alarm on container memory well below the limit.
What happened
On Sep 1, 2026 a single 14,078,689-byte PDF sent as a file part took a production task from about 640 MB to its 2 GB limit in about 40 seconds. The out-of-memory kill failed about 100 requests, 94 of them load balancer 502 errors (GitHub issue, Sep 1, 2026). The cause was the SDK's per-byte base64 encoding of file parts. Maintainers reproduced it on AI SDK 5 and 6 and closed the issue with a fix on Sep 8, 2026.
“On 2026-09-01 this OOM-killed one of our production API tasks (AWS Fargate, 2GB memory limit) in ~40 seconds: a single 14,078,689-byte PDF sent as a Uint8Array file part. The task went from a flat ~640MB to a hard kill at the container limit inside one minute, taking ~100 requests with it (94× ALB 502s) before ECS replaced it.”
safaiyeh
- Public engineering write-up (GitHub issue)
- Team metrics reported
Deployment details
A production API on AWS Fargate with a 2 GB memory limit that sends PDFs to models as AI SDK file parts. The issue names no company. Chappie is the company on the reporter's GitHub profile.
- Public engineering write-up (GitHub issue)
- Team metrics reported
Team
One production API team, unnamed in the issue
Workload
A production API that sends PDFs to models as file parts
Product version
AI SDK file parts; maintainers reproduced the fault on versions 5 and 6
Measurement
About 100 failed requests, 94 of them 502 errors, in one out-of-memory kill

Armin RonacherNov 21, 2025Armin Ronacher's Thoughts and WritingsArmin Ronacher's team built its agent on the AI SDK's provider layer under its own loop, and he said he would not make that choice again after Anthropic's web search tool kept breaking message history.
Armin RonacherFounderEarendil
An agent Ronacher's team built, using only the AI SDK's provider abstractions under an agent loop the team drives itself. The post does not name the company; Ronacher states elsewhere that he founded Earendil.
Armin Ronacher's Thoughts and Writings, published Nov 21, 2025Read the write-up (opens Armin Ronacher's Thoughts and Writings in a new tab)
Observation
Armin Ronacher's team built its agent on the AI SDK's provider layer under its own loop, and he said he would not make that choice again after Anthropic's web search tool kept breaking message history.
Takeaway
Test the server tools and caching of your main provider through the AI SDK before you commit, and keep the loop in your own code so the layer under it can change.
What happened
In a post of Nov 21, 2025, Ronacher wrote that the team adopted the AI SDK only for its provider abstractions, drove the agent loop itself, and would not make that choice again. Provider-side tools did not unify well across models. Anthropic's web search tool routinely broke the message history through the SDK, for a reason the team had not found, and cache management was easier on Anthropic's own SDK.
“The choice we made a while back was to adopt the Vercel AI SDK but only the provider abstractions, and to basically drive the agent loop ourselves. At this point we would not make that choice again.”
Armin Ronacher
Public practitioner write-up
No team metrics reported
Deployment details
An agent Ronacher's team built, using only the AI SDK's provider abstractions under an agent loop the team drives itself. The post does not name the company; Ronacher states elsewhere that he founded Earendil.
Public practitioner write-up
No team metrics reported
Team
Ronacher's agent team
Workload
An agent that calls provider-side tools such as Anthropic web search
Product version
AI SDK provider abstractions only, with the team's own agent loop

Switched to or from it2

Felipe CadavidSoftware engineerBrainGrid
BrainGrid's agent platform for turning product ideas into specs. 14 tools. Anthropic models.
BrainGrid blog, published Jul 24, 2025Read the write-up (opens BrainGrid blog in a new tab)
Observation
BrainGrid moved 14 agent tools from AI SDK 4 to the 5 beta in a couple of days, and support tickets about a frozen UI dropped to zero.
Takeaway
Upgrade for a user-visible reason, pin the exact beta, and treat the stricter types as a bug hunt: they surfaced 7 bugs the old SDK accepted.
What happened
Users saw a blinking cursor while BrainGrid's agent composed long tool inputs, so the team moved from AI SDK 4.3.16 to 5.0.0-beta.25 for tool-input streaming and provider options. Every tool definition needed renamed fields, the stricter message types broke a token counter that assumed string content, and the new stop conditions changed how long agents ran. After the migration, tool cards appear as soon as the agent starts a tool, and Anthropic cache control covers the tool definitions.
“We migrated BrainGrid's core AI infrastructure from AI SDK v4 to v5. Here's how it went.”
Felipe Cadavid
- Public engineering write-up
- Team metrics reported
- Version change stated
Deployment details
BrainGrid's agent platform for turning product ideas into specs. 14 tools. Anthropic models.
- Public engineering write-up
- Team metrics reported
- Version change stated
Team
BrainGrid engineering
Duration
A couple of days of migration work
Workload
Agent system with streaming tool calls
Product version
AI SDK 4.3.16 to 5.0.0-beta.25
Model
Anthropic Claude, with cache control on tool definitions
Measurement
Frozen-UI tickets 1 to 3 a week down to 0; tool execution 17% faster; tool-definition API cost down 7%; 7 type bugs caught

SerafimKorablevApr 7, 2026Hacker News21st.dev migrated its multi-tenant agent backend from the AI SDK to the Claude Code SDK for a complete agent loop, then rebuilt sandboxing, sessions and approvals around it.
SerafimKorablevAgent infrastructure builder21st.dev
21st.dev's multi-tenant web backend that runs AI agents for many users.
Hacker News, published Apr 7, 2026Read the write-up (opens Hacker News in a new tab)
Observation
21st.dev migrated its multi-tenant agent backend from the AI SDK to the Claude Code SDK for a complete agent loop, then rebuilt sandboxing, sessions and approvals around it.
Takeaway
Neither SDK gives a web product a multi-tenant agent runtime. Budget for sandboxes, sessions and approval screens whichever loop you pick.
What happened
In a Hacker News comment of Apr 7, 2026, the builder wrote that the team started on the AI SDK and migrated to the Claude Code SDK to get the full agent loop out of the box. That SDK turned out to be built for a single user on a command line, so the team built per-session E2B sandboxes, credential isolation, streaming, persistent threads, cost tracking and session replay itself, and rebuilt plan mode and approval prompts as React components.
“…we went through this exact thing. started on vercel ai sdk, migrated to claude code sdk hoping to get the full agent loop out of the box. same conclusion - it's built for single-user cli, not multi-tenant web backends.”
SerafimKorablev
Public practitioner write-up (Hacker News comment)
No team metrics reported
Deployment details
21st.dev's multi-tenant web backend that runs AI agents for many users.
Public practitioner write-up (Hacker News comment)
No team metrics reported
Team
21st.dev engineering
Workload
Agent sessions for many web users on one backend
Product version
AI SDK, then the Claude Code SDK

Adopted it3

Thibault Le Ouay DucasseEngineer and author of the postopenstatus
openstatus's Slack app, where the agent reads status pages and drafts incident status reports.
openstatus blog, published Feb 27, 2026Read the write-up (opens openstatus blog in a new tab)
Observation
openstatus runs its production Slack agent on the AI SDK with Claude Sonnet and six tools, and every change the agent proposes waits for a person to confirm it.
Takeaway
Have tools that change data return drafts, and put the confirmation step where the team already works.
What happened
openstatus built the agent on the AI SDK with Claude Sonnet and six tools (openstatus blog, Feb 27, 2026). Tools that would change data return a draft, and a person confirms each status report in Slack before it goes out.
“We use the Vercel AI SDK with Claude Sonnet. The agent has 6 tools:…”
Thibault Le Ouay Ducasse
- Public engineering write-up
- No team metrics reported
- Model stated
Deployment details
openstatus's Slack app, where the agent reads status pages and drafts incident status reports.
- Public engineering write-up
- No team metrics reported
- Model stated
Team
openstatus engineering
Workload
Status page reads and incident status report drafts in Slack

XataApr 2, 2025Xata blogXata tests its open-source Postgres agent with evals written as Vitest tests, saving each AI SDK response object so a failed case can be read step by step.
Richard GillAuthor of the postXata
Xata Agent, an open-source agent that diagnoses and tunes PostgreSQL databases. Several prompts and tool calls. Tool calling across different models.
Xata blog, published Apr 2, 2025Read the write-up (opens Xata blog in a new tab)
Observation
Xata tests its open-source Postgres agent with evals written as Vitest tests, saving each AI SDK response object so a failed case can be read step by step.
Takeaway
Keep the SDK's full response object as the eval artifact, and run evals in the test runner the team already uses.
What happened
Xata wanted to know whether its agent still worked after a prompt change or a model switch, so it wrote evals as Vitest tests. Vitest's global setup gives each eval run an ID, and each test case gets a folder named from its describe and test names. Cases run in parallel with describe.concurrent, which needs a local copy of expect so each result lands in the right folder. Each folder holds the AI SDK response object, with the final text, every tool call and intermediate step, and the system and user prompts. A Vitest reporter combines the results, and an internal UI shows each run so an engineer can see where a case went wrong.
“To make sure that Xata Agent still works well after modifying a prompt or switching LLM models we decided to test it with an Eval. Here, we'll explain how we used Vercel's AI SDK and Vitest to build an Eval in TypeScript.”
Richard Gill
Public engineering write-up
No team metrics reported
Deployment details
Xata Agent, an open-source agent that diagnoses and tunes PostgreSQL databases. Several prompts and tool calls. Tool calling across different models.
Public engineering write-up
No team metrics reported
Team
Xata engineering team building Xata Agent
Workload
Evals for Xata Agent's PostgreSQL diagnosis prompts and tool calls
Product version
AI SDK with Vitest
Pedro TeixeiraFeb 13, 2026metaduck.com (Pedro Teixeira)Helpmaton streams its production agents from AWS Lambda through the AI SDK, with a credit system and per-plan throttling built around it.
Pedro TeixeiraBuilder of HelpmatonDjini Labs
Helpmaton runs its production agents on AWS Lambda Function URLs, deployed with Architect.
metaduck.com (Pedro Teixeira), published Feb 13, 2026Read the write-up (opens metaduck.com (Pedro Teixeira) in a new tab)
Observation
Helpmaton streams its production agents from AWS Lambda through the AI SDK, with a credit system and per-plan throttling built around it.
Takeaway
Put cost limits and rate limits in front of the SDK from the first release; it streams and calls tools, and leaves spending and abuse to you.
What happened
In a post of Feb 13, 2026, Teixeira describes keeping Lambda Function URLs, the AI SDK and Architect as the foundation. Around them the team added streaming over several paths with the secret in the path, a credit system that protects against runaway cost, and throttling by plan on the REST API: 100 requests a second on Free and 2,000 on Pro.
“TL;DR: We kept the same foundation (Lambda Function URLs, AI SDK, Architect) and added production-ready patterns: multi-path streaming with secret-in-path auth, a credit system for cost protection, and subscription-based throttling for our REST API.”
Pedro Teixeira
Public practitioner write-up
Team metrics reported
Deployment details
Helpmaton runs its production agents on AWS Lambda Function URLs, deployed with Architect.
Public practitioner write-up
Team metrics reported
Team
Helpmaton engineering
Workload
Production agents streaming to web and REST clients
Product version
AI SDK streamText on the server and useChat in the browser
Measurement
REST throttling at 100 requests a second on Free and 2,000 on Pro

Failures, incidents and rollbacks2

In a production incident on an API Jason Safaiyeh runs, one 14 MB PDF sent through the AI SDK killed a 2 GB container in about 40 seconds and took about 100 requests with it.

Armin Ronacher's team built its agent on the AI SDK's provider layer under its own loop, and he said he would not make that choice again after Anthropic's web search tool kept breaking message history.

Switched to or from it2

BrainGrid moved 14 agent tools from AI SDK 4 to the 5 beta in a couple of days, and support tickets about a frozen UI dropped to zero.

21st.dev migrated its multi-tenant agent backend from the AI SDK to the Claude Code SDK for a complete agent loop, then rebuilt sandboxing, sessions and approvals around it.

Adopted it3

openstatus runs its production Slack agent on the AI SDK with Claude Sonnet and six tools, and every change the agent proposes waits for a person to confirm it.

Xata tests its open-source Postgres agent with evals written as Vitest tests, saving each AI SDK response object so a failed case can be read step by step.

Helpmaton streams its production agents from AWS Lambda through the AI SDK, with a credit system and per-plan throttling built around it.