@tobi You made this look too fun so I tried the same exercise with Kafka. Single rust binary, S3, one CAS manifest per partition https://t.co/fKeQFuRdZN
@vmg@tobi Before someone says it: no, this isn’t a Kafka replacement lol.
One partition per topic. No groups, no transactions.
But stock Kafka clients work. The test kills the process, starts a clean one, and reads the log back from S3.
Cursor’s “Git at any scale” by @vmg and @tobi shipping walgit over a weekend sent me down a rabbit hole.
If S3 can be the WAL for Git, why not Kafka?
So obviously I had to try it. Walstream: a tiny Kafka-compatible broker with all durable state in S3.
https://t.co/fKeQFuRdZN
@vmg@tobi The whole write path:
1. PUT a record batch
2. CAS the partition manifest
3. ACK
Lose the CAS? Read again and retry. Kill the broker? Start another one. The log is still there.
I built Lurkline: a Slack CLI + MCP that works when your workspace bans MCPs/connectors.
No Slack app. No bot. No OAuth flow. Just your existing browser session.
It also handles Slack markup without destroying it.
https://t.co/TEbxsT1Tp9
The core rule:
Search broad, read narrow.
Search filenames/snippets first. Read only the likely notes. Stop when the question is answered.
Do not dump the whole brain into every prompt.
Most AI agent memory systems are backwards.
I tried the plugin route. The better pattern was simpler:
memory = tiny bootloader
brain = searchable Markdown
session = scratchpad
hygiene = demotion loop
Memory is not the brain.
Repo + templates here:
https://t.co/1kGUgObW4m
The intended use is literal: give the Markdown writeups to your agent and ask it to implement the memory-as-bootloader + Markdown-brain pattern for your own setup.
The practical prompt is basically:
Read this repo and implement the pattern for my agent setup.
Keep startup memory tiny. Build/use a durable Markdown brain. Add templates for decisions, projects, research, and operations. Search broad, read narrow. Verify writes.
Repo with the implementation recipe + templates:
https://t.co/1kGUgObW4m
This is meant to be agent-readable: give it to your own agent and ask it to adapt the memory-as-bootloader + Markdown-brain pattern to your setup.
Great to see our 3 rd party network continue to expand!
Took insp. from Amex, back in the 90s had limited acceptance, today not far from par with VISA
Key is strong distribution partners like Google here and offer payment methods for every vertical
https://t.co/8QHVjq7NDD
Just dropped our Agentic Product Protocol: an open standard that makes 100M+ products instantly discoverable by AI agents.
-400M prices across 12 markets
-Real-time prices & availability
Before agents can buy, they need to know what exists. Now they do.
Identity thieves have a card in my name with tens of thousands of £££ in fraud.
@AmericanExpress@AskAmexUK just told me they can't help because "fraud department closes at 6pm".
So criminals get 24/7 access to commit fraud but victims have to wait for business hours?
@g_bonfiglio The difference with systems resiliency is that the sudden loss of _only_ 30% of capacity is likely a catastrophic event, you can't really plan for your company to be able to run with 2/3 of cash available.