CodeKarate

@smthomas3

[Shane Thomas @smthomas3]

"you're not really drowning in PRs... you're drowning in bad PRs"

@dexhorthy came on Agents Hour to talk about AI slop and ended up drawing the whole software factory on a whiteboard

also learn why the models produce slop by default

[Shane Thomas @smthomas3]

https://t.co/BPkhdkOigs

[Shane Thomas @smthomas3]

Overheard on a call today:

"frankly I was with my Python hat on looking at other frameworks that I was more familiar with, but also kind of realizing that I had a room full of full stack TypeScript developers and for all of the reasons that we would never use Python in production at [FAANG company], using Python was probably not a good idea for this either"

[Shane Thomas @smthomas3]

deploy your agents to mastra cloud

you can just deploy things

[Shane Thomas @smthomas3]

almost there...

[Shane Thomas @smthomas3]

Excited to announce we just released a @mastra "course"

But it's unlike any course you have taken before. Take the course right from your IDE and it walks you through writing the code.

Here is how it works...

[Shane Thomas @smthomas3]

AI Agents Hour, May 15th https://t.co/RkrAYsYgwc

[Shane Thomas @smthomas3]

customer switching from langchain

"all we wanted in ai was in mastra"

[Shane Thomas @smthomas3]

should you put ai engineer as your job title on linkedin?

w/ @abhiaiyer

[Shane Thomas @smthomas3]

multi-agent systems are the future of ai applications (with @abhiaiyer)

[Shane Thomas @smthomas3]

making a list, checking it twice, my agent will buy her something quite nice

never going to gift shop the same again

[Shane Thomas @smthomas3]

Ready to up your "vibe coding" game? Join @abhiaiyer and me at our Thursday workshop to learn tips/tricks we are using at @mastra (link below)

- Vibe coding isn't just letting AI write your code - it's about creating a collaborative workflow with AI agents
- The key difference: "knowing what you want" vs "exploring something new" requires different prompting strategies
- Learn how to create detailed checklists for AI to follow when you know the goal
- Quick tip: Ask the AI to research first when working on unfamiliar tasks
- Join Thursday's workshop to share strategies for getting better results from tools like Cursor and Windsurf

[Shane Thomas @smthomas3]

Want to learn how to build voice agents? (with @zammitjames and @dan_gauci from Roark)

- The 3 critical components of voice agents
- The future is speech-to-speech technology that combines all 3 layers into one
- OpenAI's realtime API and Gemini are early players, but production adoption is still limited
- Join our Voice Agents Workshop TOMORROW (April 17th) to learn how to build, monitor, and test a voice agent

[Shane Thomas @smthomas3]

MCP makes it easy to connect a @mastra agent to all your tools.

I'm teaching a workshop on all things MCP. (link below)

If you can't wait until Thursday, here's a simple example you can build in 5 minutes to connect an agent to a Google Sheet

[Shane Thomas @smthomas3]

AI agents that don't learn from their mistakes are doomed to fail (with @viewerisland from @Gulp_AI)

- Gulp's API analyzes agent interactions to improve performance on similar tasks
- Unlike RAG or memory solutions, Gulp acts as a proxy layer that rewrites prompts to be more detailed and directed
- Traditional memory retrieval suffers from "noise" - returning too much irrelevant information
- Real results: GPT-4o-mini with Gulp API achieved nearly 60% performance improvement over standard GPT-4o on web arena benchmarks
- The key insight: Bad input equals bad output - focusing on prompt quality rather than fancy infrastructure

[Shane Thomas @smthomas3]

The voice AI opportunity is here—but solving the last 20% is where the real challenge lies (with @zammitjames and @dan_gauci from Roark AI)

- Voice AI is becoming the new frontier with companies like Roark building "datadog for voice AI" to help developers debug their agents
- Popular use cases include lead generation (outbound sales calls) and appointment scheduling for industries like dental clinics and car dealerships
- Getting to 80% functionality is easy, but the last 20% is where voice agents struggle significantly
- Major issues: 30% of transcriptions are incorrect (some companies hire humans to manually transcribe), conversations deteriorate over time, and tool calling is notoriously inaccurate
- Voice agents have trouble recovering when conversations go off-script, leading to hallucinations and poor user experiences

[Shane Thomas @smthomas3]

The story of how my @mastra demo video accidentally got 27k views on YouTube

🧵 👇 For the full story and some product demo tips that worked for us

[Shane Thomas @smthomas3]

Wondering why your timeline is suddenly flooded with AI-generated visuals? Here's what just happened with GPT-4o.

- OpenAI's GPT-4o image generation capabilities are creating a new visual style taking over social media
- The new model excels at consistent characters, incorporating text, and following specific prompts
- Even with modest design and editing skills, these tools let you create professional content in under 2 hours
- This isn't about AI taking jobs - it's about AI users outperforming everyone else
- Developers: API access for these capabilities is coming soon

[Shane Thomas @smthomas3]

You just can't fully replace the feeling of getting out the whiteboard and drawing an architecture diagram. The only problem, it eventually gets erased and lives on as a photo that no-one looks at again...

I used @mastra to turn whiteboard sessions into @excalidraw diagrams.

[Shane Thomas @smthomas3]

APIs were built for humans - but LLMs need something different (with @Life_of_Y_ from @wildcard_ai)

- OpenAPI specs have decades of experience behind them, but weren't designed with AI agents in mind
- Traditional APIs use imperative commands that LLMs struggle to interpret without proper context
- The problem: LLMs don't understand concepts like pagination or checkout flows without explicit teaching
- Agents.JSON offers "flows" - deterministic sequences of API calls that reduce the need for LLM reasoning
- Pre-planning API interactions could save compute across millions of AI agents instead of each one planning at inference time