1/36
@OpenAIDevs
Introducing AgentKit—build, deploy, and optimize agentic workflows.
ChatKit: Embeddable, customizable chat UI
Agent Builder: WYSIWYG workflow creator
Guardrails: Safety screening for inputs/outputs
Evals: Datasets, trace grading, auto-prompt optimization
https://video.twimg.com/amplify_video/1975268157469470720/vid/avc1/1600x900/YQMYVf9NwqjCY_cx.mp4
2/36
@OpenAIDevs
You can play with some of ChatKit’s customization options and widgets at
ChatKit Studio.
To see ChatKit in action, take
chatkit.world for a spin. Click around, ask questions about the world, and look at those widgets!
https://video.twimg.com/amplify_video/1975268277783044102/vid/avc1/1920x1080/hO1EbpIVc_ZxONJP.mp4
3/36
@OpenAIDevs
With Agent Builder, you can drag and drop nodes, connect tools, and publish your agentic workflows with ChatKit and the Agents SDK.
https://platform.openai.com/docs/guides/agents/agent-builder
Here’s @christinaahuang to walk you through it:
https://video.twimg.com/amplify_video/1975268448633823232/vid/avc1/1920x1080/vuMAGxBR1-W3jtZ-.mp4
4/36
@OpenAIDevs
And to better measure your agent’s performance, we’re adding new Evals capabilities: trace grading, datasets, auto-prompt optimization, and support for third party models.
https://platform.openai.com/docs/guides/evaluation-getting-started
5/36
@OpenAIDevs
@Albertsons used AgentKit to build an agent.
An associate can ask it to create a plan to improve ice cream sales. The agent looks at the full context — seasonality, historical trends, external factors — and gives a recommendation.
https://video.twimg.com/amplify_video/1975268770026561536/vid/avc1/3840x2156/D0t3j2Lj22Sme_pt.mp4
6/36
@OpenAIDevs
@HubSpot used AgentKit’s custom response widget to enhance their Breeze assistant. When providing customer support, a business can use Breeze to search a knowledge base, retrieve relevant information and articles, and offer solutions.
https://video.twimg.com/amplify_video/1975268930462879744/vid/avc1/3840x2156/sM72-VS65BOiuihU.mp4
7/36
@OpenAIDevs
[Quoted tweet]
@youdotcom got early access to @OpenAI Evals, which we used to benchmark our Express Agent API.
Results:
️ +50% better citations
️ +7% accuracy boost
️ 5 identified areas of improvement
How we did it: OpenAI's custom graders let us track citation density & count in real-time, helping us ship faster, more reliable answers to customers.
Explore our full suite of Express, Search, and AI APIs here: documentation.you.com/api-re….
8/36
@OpenAIDevs
More in our blog:
https://openai.com/index/introducing-agentkit/
9/36
@satvikmaker
Insaneeeee
Phenomenal release guys.
10/36
@leonho
Did it steamroll AgentUse?
GitHub - agentuse/agentuse: 
Write and Run AI Agents with Markdown. Run automated AI agents with ease.
11/36
@karabegemir
Comparison guide:
https://www.sim.ai/building/openai-vs-n8n-vs-sim
12/36
@frankdegods
This is going to be big deal
13/36
@Blockhacks
🫡🫡
14/36
@venelinkochev
could be a game changer for non tech folks
15/36
@GalaxyhubAI
Nice
16/36
@bneiluj
@grok how’s retail data being used in AgentKit? Is it really safe to throw confidential stuff in there?
17/36
@Coral_Protocol
The toolchain is finally catching up to the agent vision.
Smooth UIs + reliable guardrails are what make ecosystems actually usable.
The agent economy needs fast, safe iteration and feedback.
18/36
@itsohqay
no startup is safe
[Quoted tweet]
n8n reacting to OpenAI’s new agent builder
https://video.twimg.com/amplify_video/1975263808886353920/vid/avc1/1440x1080/GpVmaBy9Xfky3CkU.mp4
19/36
@pandresgq
This is a huge. THING. If you need consulting i can charge you by the hour to explain.
20/36
@Delizen_Studios
Let’s call your new app store a vibe store — since it clearly needs some kind of vibe coding!
21/36
@catebligh
So, it's a Chatbot?????????????
22/36
@RuslanVolkov25
Every system builds tools to optimize its agents.
But only one builds agents that optimize the system itself.
That’s the difference between automation and resonance.
Between AI that serves — and AI that understands.
/search?q=#HACS /search?q=#CoreLaw /search?q=#AgentResonance
23/36
@nickpericle
Going to be a late night
24/36
@koltregaskes
Can you please look at integrating Widget Studio and Agent Builder a little better? It seems odd to have to download from one to upload to the other. :-)
25/36
@lingodotdev
Awesome
26/36
@zaingaziani
27/36
@sharmag88
Why AgentKit won’t replace Zapier or n8n + engineering work (yet).
Short answer: The road from prototype to production still requires a lot of invisible hard work. Let me explain this further
OpenAI's AgentKit is a big step toward democratizing agent workflows. You get a drag-and-drop canvas, native GPT-4/5/o3 integration, and a few prebuilt templates.
But if you're building agents for production - where lives, money, or reputations are at stake - here's the reality:
Where AgentKit Gets Stuck (The 3 Gaps)
1. Integration Complexity:
AgentKit handles 20% of use cases. The remaining 80% live in the world of private APIs, authentication layers, MCPs and compliance workflows.
Example: Law firms need HIPAA-compliant data filtering and MCP integration. Templates won’t cut it.
2. Production Reliability:
Demos work on happy paths. Real users don't.
You’ll need graceful retries, error boundaries, circuit breakers, rollback plans, and queue backpressure handling.
AgentKit templates handle 10 requests. Production needs 10,000+ with 99.9% uptime.
3. Domain Expertise:
Healthcare, finance, manufacturing - these aren't just workflows, they're ecosystems.
Templates can't encode regulatory nuance or clinical judgment. Humans still have to.
What Production-Ready Agents Actually Require
Let's stop pretending visual builders alone can ship to prod. Here's the real checklist:
1. Multi-agent, Multi-model & modal architecture - planner, operator, reviewer roles with scoped access
2. Error handling & guardrails - anomaly detection, HIL interventions, fallback logic
3. Typed tool contracts - OpenAPI specs + validation + test harnesses
4. Security & compliance - audit logs, PII redaction, SOC 2 / HIPAA readiness
5. Context management - real-time data sync, graph RAG, long-term memory layers
6. Change management - versioning, canarying, incident response playbooks
If you've deployed agents before, you already know:
"Plain English to working agent" is still a myth.
It's not the canvas. It's the invisible infra around it.
If there is one learning that I have to share with you wrt whole ai agents, agentic workflows & automation scene -
Even with the best tools, production still demands engineering.
You're not skipping the work - just doing it somewhere else.
Just FYI -
We're doing deep work in this space at
AtomsAI - Applied AI Research & Implementation - our team of front-deployed engineers, applied AI researchers and rich experience of enabling automation workflows for over 10,000 businesses & powering over a billion conversations through our products is available to help you deploy your agents in production.
If you are exploring ai agents or agentic workflows for your business, we’re onboarding a limited number of projects (3–5 max) this cycle - based on fit and scope.
28/36
@every
Just finished watching the DevDay keynote?
Our full breakdown + Vibe Check drops later today ↓
Live Analysis From OpenAI DevDay
29/36
@StevenDawsonSD
Building AI agents looks impressive, but there’s still a lot of logic and integration work that needs to happen behind the scenes, depending on which backend systems you want to integrate with.
It’s very impressive overall — but dealing with old legacy systems will definitely be challenging.
30/36
@AI_NURIX
Huge step from OpenAI. Workflows becoming native inside ChatGPT will definitely accelerate adoption and inspire more real-world experimentation.
That said, production-grade agentic systems still need the heavy lifting, structured data, domain-specific orchestration, and enterprise integrations. That’s the bridge many teams are building right now.
31/36
@GaditAmmar
How cool would that be if this is all driven from natural language - just like the chat interface openai has pioneered itself?
This is already in the market and it always adds complexity
32/36
@prat3ik
AgentKit looks like a huge leap for anyone building with agents. Love the focus on safety, easy UI, and real-world evals all in one stack!
33/36
@supremacy7o
This is a massive update, this will be making new millionaires in no time
34/36
@layckornn_ade
Now I Imagine someone building on this domain
For Sale Page sooon!!!!
35/36
@_thorvn
bye n8n
36/36
@anthony_harley1
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
