Sebastian Korfmann
LightNews LightNews
skorfmann.com
Sebastian Korfmann
@skorfmann.com
1.4K followers 750 following 1.5K posts
Exploring MCP and agentic coding, both personally and for the enterprise context.

Based in #hamburg - https://skorfmann.com
Posts Replies Media Videos
Reposted by Sebastian Korfmann
simonwillison.net
Sprites.dev by @fly.io is a very cool new thing: it solves two of my pet problems at once, developer sandbox environments for coding agents and a JSON API for executing untrusted code

I wrote more here: simonwillison.net/2026/Jan/9/s...
Fly’s new Sprites.dev addresses both developer sandboxes and API sandboxes at the same time
New from Fly.io today: Sprites.dev. Here’s their blog post and YouTube demo. It’s an interesting new product that’s quite difficult to explain—Fly call it “Stateful sandbox environments with checkpoin...
simonwillison.net
January 10, 2026 at 12:32 AM
skorfmann.com
❄️🌨️🥳
January 8, 2026 at 6:50 AM
skorfmann.com
210 RSVPs and we’re at capacity for our next Agentic Coding meetup in Hamburg – the “Claude Code special” edition 🔥

Incredible numbers by Hamburg standards. Something is shifting – and it’s shifting fast.

agentic.hamburg​​​​​​​​​​​​​​​​
Agentic Hamburg - The Coding Meetup for Hamburg
Regular meetups where developers share AI workflows, demo tools, figure out what actually works, and inspire each other.
agentic.hamburg
January 7, 2026 at 6:55 AM
skorfmann.com
Cross region transfer cost - where the actual money is made for cloud providers 💸
January 1, 2026 at 12:23 AM
skorfmann.com
December 29, 2025 at 8:15 PM
skorfmann.com
> This PR represents excellent engineering work
December 28, 2025 at 12:16 AM
Reposted by Sebastian Korfmann
timkellogg.me
Ralph: a Claude Code plugin that uses stop hooks to keep Claude chugging. if it tries to stop, ralph feeds the same prompt and it keeps going

apparently Anthropic employees keep Claude going for hours and even days like this

github.com/anthropics/c...
claude-plugins-official/plugins/ralph-wiggum at main · anthropics/claude-plugins-official
Anthropic-managed directory of high quality Claude Code Plugins. - anthropics/claude-plugins-official
github.com
December 27, 2025 at 4:08 PM
skorfmann.com
Getting more and more used to worktrees - running three sessions in parallel right now #claude
December 27, 2025 at 3:18 PM
Reposted by Sebastian Korfmann
hmans.dev
👀 Some words about Chatto, my BIG project. It's a group chat app that

- is incredibly easy to self-host
- has very modest hardware requirements
- scales really well
- apache-2.0 license
- hosted platform launching in 2026

Looking forward to writing much more about this soon.

Let's do this. 🚀
Introducing Chatto
I'm building an Open Source group chat application that's simple to self-host, familiar to use, and easy to customize.
www.hmans.dev
December 26, 2025 at 9:50 PM
skorfmann.com
Remember Rails Rumble? 48 hours, team of 3-5, building an app from zero to production.

Today, a single dev with Claude Code could easily compete.
December 21, 2025 at 9:01 PM
skorfmann.com
Posting this through Claude Code via the Claude in Chrome extension! 🤖✨ (Had to upload the image manually though 😅)
December 18, 2025 at 4:47 PM
Reposted by Sebastian Korfmann
timkellogg.me
Gemini 3 Flash

Basically Gemini 2.6 Pro but with Flash pricing & speed

blog.google/technology/d...
A large comparison table showing AI model pricing and benchmark performance across many tasks. Column headers (left to right): Benchmark | Description | Gemini 3 Flash (Thinking) | Gemini 3 Pro (Thinking) | Gemini 2.5 Flash (Thinking) | Gemini 2.5 Pro (Thinking) | Claude Sonnet 4.5 (Thinking) | GPT-5.2 (Extra high) | Grok 4.1 Fast (Reasoning) Pricing rows (top): • Input price ($/M tokens): Gemini 3 Flash $0.50 | Gemini 3 Pro $2.00 | Gemini 2.5 Flash $0.30 | Gemini 2.5 Pro $1.25 | Claude Sonnet 4.5 $3.00 | GPT-5.2 $1.75 | Grok 4.1 $0.20 • Output price ($/M tokens): Gemini 3 Flash $3.00 | Gemini 3 Pro $12.00 | Gemini 2.5 Flash $2.50 | Gemini 2.5 Pro $10.00 | Claude Sonnet 4.5 $15.00 | GPT-5.2 $14.00 | Grok 4.1 $0.50 Selected benchmark rows (scores shown as percentages unless noted): • Humanity’s Last Exam: Gemini 3 Flash 33.7% / 43.5%, Gemini 3 Pro 37.5% / 45.8%, Gemini 2.5 Flash 11.0%, Gemini 2.5 Pro 21.6%, Claude 13.7%, GPT-5.2 34.5% / 45.5%, Grok 17.6% • ARC-AGI-2: Gemini 3 Flash 33.6%, Gemini 3 Pro 31.1%, Gemini 2.5 Flash 2.5%, Gemini 2.5 Pro 4.9%, Claude 13.6%, GPT-5.2 52.9% • GPQA Diamond: Gemini 3 Flash 90.4%, Gemini 3 Pro 91.9%, Gemini 2.5 Flash 82.8%, Gemini 2.5 Pro 86.4%, Claude 83.4%, GPT-5.2 92.4%, Grok 84.3% • AIME 2025: Gemini 3 Flash 95.2% / 99.7%, Gemini 3 Pro 95.0% / 100%, Gemini 2.5 Flash 72.0% / 75.7%, Gemini 2.5 Pro 88.0%, Claude 87.0% / 100%, GPT-5.2 100%, Grok 91.9% • MMMU-Pro: Gemini 3 Flash 81.2%, Gemini 3 Pro 81.0%, Gemini 2.5 Flash 66.7%, Gemini 2.5 Pro 68.0%, Claude 68.0%, GPT-5.2 79.5%, Grok 63.0% • ScreenSpot-Pro: Gemini 3 Flash 69.1%, Gemini 3 Pro 72.7%, Gemini 2.5 Flash 3.9%, Gemini 2.5 Pro 11.4%, Claude 36.2%, GPT-5.2 86.3% (with python) • CharXiv Reasoning: Gemini 3 Flash 80.3%, Gemini 3 Pro 81.4%, Gemini 2.5 Flash 63.7%, Gemini 2.5 Pro 69.6%, Claude 68.5%, GPT-5.2 82.1% • OmniDocBench 1.5 (lower is better): Gemini 3 Flash 0.121, Gemini 3 Pro 0.115, Gemini 2.5 Flash 0.154, Gemini 2.5 Pro 0.145, Claude 0.145, GPT-5.2 0.143 …
December 17, 2025 at 5:16 PM
skorfmann.com
would be great if claude code could watch for github actions / pr comments and reacts automatically (on my local machine) and only pings if decisions are required
December 17, 2025 at 4:01 PM
skorfmann.com
time to sleep, I guess status.claude.com
December 14, 2025 at 9:53 PM
skorfmann.com
Just used Beans for Claude Code for a smallish refactoring - and I quite like it. Simple, markdown files, nice TUI to see what's going on and potentially even edit stuff. Claude Code and the TUI on split screen.
December 14, 2025 at 8:35 PM
Reposted by Sebastian Korfmann
ingwarpero.bsky.social
Wow! Vor 24h habe ich mit viel Herzklopfen Cycling Cities veröffentlicht. Jetzt kann ich sagen: Ich bin total überwältigt von den vielen lieben Nachrichten & so erleichtert, dass der Film so positiv ankommt. Vielen lieben Dank & ich wünsche weiter ganz viel Spaß beim Film🍿

youtu.be/R5jQc1jXLAU?...
December 14, 2025 at 8:17 AM
skorfmann.com
blog.joemag.dev/2025/10/the-...
The New Calculus of AI-based Coding
A collection of posts by Joe Magerramov
blog.joemag.dev
December 14, 2025 at 7:17 AM
skorfmann.com
Sounds like an actually useful ai case in a browser www.theverge.com/tech/842000/...
Google is building an experimental new browser and a new kind of web app
Meet the GenTab, nothing like the plain ol’ browser tab.
www.theverge.com
December 13, 2025 at 7:45 AM
Reposted by Sebastian Korfmann
hmans.dev
It's so much fun to leave `beans tui` running in a separate window. It'll automatically pick up changes in the filesystem and update itself accordingly.
The terminal UI built into Beans, my CLI-based issue tracker for humans and robots.
December 12, 2025 at 10:52 PM
skorfmann.com
Deep research api blog.google/technology/d...
Build with Gemini Deep Research
We have reimagined Gemini Deep Research to be more powerful than ever, now accessible to developers via the new Interactions API.
blog.google
December 11, 2025 at 6:03 PM
skorfmann.com
The Agent Engine looks quite promising docs.cloud.google.com/agent-builde...
Vertex AI Agent Engine overview  |  Vertex AI Agent Builder  |  Google Cloud Documentation
Learn about Vertex AI Agent Engine, a set of services that enables developers to deploy, manage, and scale AI agents in production.
docs.cloud.google.com
December 10, 2025 at 8:08 PM
Reposted by Sebastian Korfmann
mathias.blog
Great for monorepos: you can now define instructions in .claude/rules/ that are applied based on glob patterns, e.g. paths: src/api/**/*.ts

code.claude.com/docs/en/memo...
Manage Claude's memory - Claude Code Docs
Learn how to manage Claude Code's memory across sessions with different memory locations and best practices.
code.claude.com
December 10, 2025 at 1:44 PM
Reposted by Sebastian Korfmann
hmans.dev
Beans and Claude are such a great team. Instead of whipping up a plan that will be long forgotten the very next session, just put it into a bean, and it'll stick around until you want to iterate on or implement it.
--- title: Refactor prompt command to use text/template status: todo type: task created_at: 2025-12-08T20:16:08Z updated_at: 2025-12-08T20:16:08Z --- ## Summary Refactor the `beans prompt` generation to use Go's `text/template` package instead of string builders and manual string concatenation. ## Motivation The current implementation likely uses `strings.Builder` or similar manual string construction, which can be harder to read, maintain, and modify. Using `text/template` would: - Make the output format more declarative and easier to understand - Separate the template structure from the data logic - Make it easier to customize or extend the prompt format in the future - Follow Go best practices for text generation ## Checklist - [ ] Review the current `beans prompt` implementation to understand its structure - [ ] Create a template file or embedded template string for the prompt output - [ ] Refactor the command to use `text/template` with appropriate data structures - [ ] Ensure the output remains identical to the current implementation - [ ] Add/update tests to verify the refactored code produces correct output
December 8, 2025 at 8:25 PM
Reposted by Sebastian Korfmann
hmans.dev
It's my birthday, but today *you* get the presents!

Introducing "Beans", a CLI-based, flat-file issue tracker designed for humans and their robot friends. 🫘

Track your work right in your repo, using the CLI. And if you work with coding agents, supercharge them with a full picture of your project!
December 8, 2025 at 3:36 PM
Reposted by Sebastian Korfmann
mathias.blog
Made a thing: Overcast Unwrapped. Drop your OPML export and explore your podcast listening history. Runs entirely in-browser. Free to use and copy.

mathias.blog/tools/overcast
December 5, 2025 at 3:09 PM