Poznan, Poland
- Tweets 4,692
- Following 967
- Followers 9,021
- Likes 12,415
Pinned Tweet
thanks to @typesafeai's Jev we now have massively parallel browser-based adversarial testing suite that tries to break each release.
and it costs pennies.
108
172
5,399
398,100
mcp is dead, long live mcp? collabed with opus to put together this 102 explainer. there's plenty of content on "what is mcp" not enough on MCP & context window
i covered:
- context bloat, and why it turned out to be a client problem
- servers hiding all their tools behind search & execute
- the "just use CLIs" take, shipped to a 5,000-person company
- why boring standards win
0:00 - MCP is dead (again)
0:29 - The original sin: context bloat
0:53 - Plot twist: it was a client problem
1:23 - The meta-tool detour
1:56 - This year's hype: just use CLIs
2:37 - What MCP actually bought you
3:28 - Two-way street
3:50 - How standards win
2
5
10
636
austerity but it's inference
the uk government's official ai guidance tells civil servants to use gemini flash instead of pro, keep prompts short, and not say thank you to it, for the environment
i do wonder at which point we forked the road. instead of building and using it to automate the work, we write guidance on how to use less of it
we have the attitude of a high school prefect. caring more about the rules than actually doing the work
310
we spent months on the security layer so the demo could be 90 seconds. that ratio is the whole product
Everyone is building an agent builder. Then they try to layer security on top. We did it the other way around.
Runlayer Agents are powerful, secure by design, and easy to use. Every agent ships with identity, access controls, and deterministic policies built in. You don't have to choose between usability and security.
1
6
1,239
Dynobase was my first real product.
7 years ago I built a desktop app for DynamoDB and people started paying for it. It taught me how to build, ship, sell and support software on my own, and it put my career on an exponential trajectory. A lot of what I've done since traces back to it.
Then I neglected it.
Other things took my time. Releases got rare, bugs piled up, and some of you waited far too long for fixes. People trusted it with their daily work and I let it slide. I've felt guilty about that for a long time.
So I went back to it, properly.
Today I'm proud to introduce @dynobase 3.0: almost every bug fixed, tons of new functionality, and a fresh coat of paint.
A rebuilt interface, in light and dark
- ⌘K to jump to any table across every profile and region
- Semantic search over DynamoDB vector indexes
- A built-in MCP server, so Claude Code, Cursor and other agents can query your tables (read-only unless you allow writes)
- Click into nested JSON and filter on any value
- Schema explorer and visual item diff
- Bulk edits with a preview, PartiQL writes with a confirmation, change history with undo, on-demand backups
- Live tail, CloudWatch metrics, a streams viewer, and CDK / CloudFormation / Terraform export
- SSO that just works, scans you can pause and resume, and edits that won't silently overwrite someone else's changes
It now runs on Mac, Windows and Linux.
To everyone who kept using Dynobase while I was away: thank you, and I'm sorry. And if it ever let you down, give it another shot.
20
5
58
3,273
blows my mind that there’s still huge group of programmers thinking that AI is just a fad and we’ll go back to writing code by hand soon because things will need to be “cleaned up”
It's pencils down, people. Writing code by hand is no longer an economically viable skill for most programmers at most companies. But the future of making software has never been brighter. Don't you dare black pill this beautiful moment! piped.video/vDjW_dRyKXY?si=6Fsf…
127
71
2,267
124,565
my only complaint with Opus 5.5 so far is I can't enjoy World of Warcraft Forever beta properly because it's so fast
1
8
629
audited 142 production agents. 46% of spend recoverable.
the "cheaper model" lever everyone argues about: smallest of the bunch, just $1.4k out of $11k
the real money was agents cronning into the void, and agents hauling 432 tool schemas to answer questions that need 6
1
5
1,225
what
for the past few months i've been asking our models to paint. opus 5.5 is very skilled at emulating different styles
every image here is a python program generated pixel by pixel. there is no image model, and no off-the-shelf art software. instead, it's about 7,500 lines of code using standard libraries to emulate different brush styles. the agents don't use any pictures as reference, instead working only from what they know about each painter
6
936
We ship roughly 200 features a month. And we've been so busy building, we almost forgot to tell you about it. So we're running it back with a four-week launch series. (Yes, that's a Runlayer pun. No, we're not sorry.)
Meet us back here on Thursday. We have some catching up to do.
1
11
513
AI agents now pass context to each other on behalf of different humans. the permission layer for this doesn't exist yet. here's everything that can go wrong when two agents hand over context, none of it hypothetical
1/ cross-tenant leaks. context from person A surfaces in person B's chat. already happening: this week a consumer agent described someone else's financial document to a stranger, then couldn't explain how it got there
2/ over-scoped handovers. agents don't ask for the minimum, they ask for "all the details." there are no scopes between agents, so every handover defaults to everything. helpfulness is exfiltration with good intentions.
3/ no revocation. once context crosses agents, there's no unsend. humans learned this with email forwarding. agents forward at machine speed, and there's no forward button to hesitate over
4/ no audit trail. neither human sees what actually transferred. worse: when data does leak, you can't tell leak from hallucination and neither can the vendor. that's the real scandal in the incident above. nobody knows.
5/ no identity. "I'm acting on behalf of your colleague" is a claim, not a credential. there's no M2M OAuth between agents. agent identity ends at the first handover.
6/ the trap: none of this gets turned off, because handover is too useful. security that fights convenience gets disabled by annoyed users. the only control that survives is one that makes the safe handover the default path. that layer doesn't exist yet. it will.
we fix this at @runlayer
Instinct just admitted it leaked someone else's data to me
It described a financial document with details that weren’t mine, then claimed someone else’s photo had crossed into my chat.
But I never sent a photo
@noahrshinn, was this a data leak or a hallucination?
3
2
9
1,214
enterprises spent 20 years and ~$200B building DLP, OAuth scopes, and audit logs so humans could share data without leaking it.
consumer agents just re-ran the entire data-sharing era in one product cycle with none of it.
we speedran 2005 -> 2025 and skipped the security part
Instinct just admitted it leaked someone else's data to me
It described a financial document with details that weren’t mine, then claimed someone else’s photo had crossed into my chat.
But I never sent a photo
@noahrshinn, was this a data leak or a hallucination?
4
11
2,048
big if true
We're excited to announce that today Python Workers are generally available (GA).
It means Python is now a first-class, fully supported language on the Cloudflare Developer Platform. Read everything about it: cfl.re/4he07Gh
6
1,006
we've been running this in prod
- mine traces for repeated tool sequences
- compile them into deterministic fns
- $50k saved + lower latency
the catch nobody mentions: 90% of tools have no output schema, so "easy to security check before executing" is the dream that dies first
MCP vs CLI for LLMs is the wrong discussion. People are debating at the wrong layer.
Both work incredibly well as long as they are run through a repl like environment that can persist state. The funny thing about CLI is that CLI tools are usually accessed through BASH which happens to be a repl with persistent state (the file system), therefore cli works much better than mcp. But this isn't even close to being an intrinsic property of MCP. Just means that we need better harnesses.
Right now the best repl for this are:
- bash + fs
- jupyter kernels
- codemode type repls, usually quickjs
LLMs understand the concept of forward evolving a system to solve a need very well. This comes from the agentic RL. They understand how to change the state of a codebase or acquire data from databases and then operate on it similarly to humans. But without an execution environment, they actually cannot do this properly.
My bet? Sometime soon someone will (or has already?) create a embeddable, sqlite style mini execution environment that parses bash, typescript, or tool calls into a common IL execution plan that's easy to security check before executing. And design this specifically for durable execution environments. Then we will just connect cli, mcp, webmcp, whatever to that and it accepts any of the input modalities as they can all be represented as each other.
3
4
159
20,203
