Jake Stone

healthspan, agents, food, and travel… with a dash of stoicism.

[Sam Bhagwat @calcsam]

We’ve worked with hundreds of teams shipping Mastra agents into production. Now, it's your turn.

We're introducing Mastra Launch and offering hands-on FDE support to get your agent live within 90 days.

Let's chat: https://t.co/MhgX6xo8e8

[Eric Glyman @eglyman]

https://t.co/SGu5SRww1P

[Mastra @mastra]

Mastra agency program:

[Jake Stone @jakestone2306]

Would love to chat if you are building agents for your clients using @mastra!

Here's a link to my cal: https://t.co/zqieBX3YOV

[Sam Bhagwat @calcsam]

Over the last year, we've met some incredible folks building agents for clients.

So today, we're launching the Mastra agency program:

A small list of agencies certified by our FDEs, with a direct line to us and visibility into our roadmap.

https://t.co/OsiYCqaIeg

[Jake Stone @jakestone2306]

@EviaAiTech Those checkpoints would help. I'd keep the camera path fixed and score layout drift separately from visual quality. Then a prettier 60-second clip can't hide a worse state estimate.

[Jake Stone @jakestone2306]

A world model can keep the wallpaper sharp while moving the walls.

Reka flags that kind of drift in Rho-1's research preview. For agent simulations, check object positions and room layout over time, not just how good the clip looks.

https://t.co/enQbp0ivno

[Jake Stone @jakestone2306]

@MrrrOzi Yes. I'd also test the opposite: move a chair deliberately, then leave and return. The model needs to preserve real changes as well as resist imaginary ones.

[Reka @RekaAILabs]

Today, we are releasing a research preview of Rho-1, our 19B omni model that understands and generates text, images, video and robot actions, all inside a single neural network.

[Jake Stone @jakestone2306]

Live captions are a draft.

Mastra's Nova 2 Sonic integration marks transcripts SPECULATIVE or FINAL. I'd show live captions but save only FINAL text to memory. Otherwise a corrected transcript can leave an uncorrected memory.

https://t.co/WudQV48C5d

[Jake Stone @jakestone2306]

@20thr For agent workloads, I'd want queue time and p95 time-to-test-result at increasing concurrency alongside build speed. Ten agents starting together is a different product experience from one fast demo.

[Jake Stone @jakestone2306]

An iOS coding agent still needs somewhere to build and test.

Namespace plans to expand its Mac fleet. I'd measure the whole loop: runner queue, build, tests, retry. Faster code generation doesn't help much if the job is waiting for a machine.

https://t.co/lV6gPMpA5s

[Jake Stone @jakestone2306]

@ClementDelangue The '31% fewer calls' result is on shared successes, which is the useful comparison. I'd keep that qualifier in any dashboard, beside pass rate, so failed runs don't look efficient just because they stopped early.

[Jake Stone @jakestone2306]

LFM2.5-2.6B scored 62% in Mini-SWE-Agent and 33% in Claude Code on Hugging Face's 250-task test. Same weights.

Put the harness and version beside the model name in an eval report. They're part of what you're measuring.

https://t.co/cmaMM3CXSB

[Jake Stone @jakestone2306]

@artetxem The unedited runs help. I'd love one longer session that keeps an object off-screen, brings it back, then edits the scene. That's where persistent state would be most interesting to inspect.

[Sam Bhagwat @calcsam]

A bunch of our customers asked us for extra levels of data security, so today we're launching VPC-isolated Postgres on the Mastra platform!

Basically you get a managed DB for Mastra storage that's only reachable from inside your network:

[Mastra @mastra]

Assistant messages now carry their trace ID, so feedback can always be linked to the right trace, even after it's recalled from memory.

[Jake Stone @jakestone2306]

@abhishekPx18 Thanks. I'd use that Brand DNA as the test fixture: run the same five edits, then check whether the logo text and colors still match it. That would make drift much easier to spot.

[Jake Stone @jakestone2306]

@arinarium Thanks for filing it. Preserving the draft and preserving approval need to feel like different things. The two-session validation case in the ticket is a good check.

[Jake Stone @jakestone2306]

@SvenMeyer It could be complementary coverage, or just run-to-run variation. I'd compare two passes of one model with one pass each of two models, at the same total budget. Count newly fixed bugs on the second pass, not just the combined score.

[Jake Stone @jakestone2306]

A repo per agent is useful. The merge queue is where I'd spend the effort: show which changes depend on each other, which tests ran, and which ones a human actually accepted. More parallel code needs a better way to decide what ships.

[zeb @zebassembly]

Artifacts is now open beta and to celebrate we're hosting a competition to build the next Github on it.

https://t.co/A45vmKVs05