Computer scientist building the infra AI needs to scale around search engines, databases, and distributed systems.
Earth
- Tweets 11,557
- Following 393
- Followers 773
- Likes 63,241
Do you maintain the code that your coding agent writes, or throw it away and have the agent write it again whenever it breaks?
ALT On a dark square card, the sentence 'However quickly your coding agent writes the code, you will have to maintain it for as long as your customers depend on it' appears in large type, with 'you will have to maintain it for as long as your customers depend on it' in blue.
10
Patroni is a Python template for building PostgreSQL clusters that keep a database available when one of its servers fails: github.com/patroni/patroni
11
Claroc's researchers show how Muse put each chat's title into the domain name of its share link, which is how those titles ended up in DNS records that they could search: blog.claroc.io/ghostshares-m…
4
If your AI agent can reach your data only by looking up an ID, it can tell a customer that you have nothing on a subject, although several of your records are about it.
The customer's question contains no ID, which means that the model has to guess one. When the guess matches no record, the agent has nothing to answer from.
Tell your coding agent to add a search tool that takes the customer's own words. Keep the lookup by ID for opening a record that the search tool returns.
This guide has the code for the search tool and the lookup by ID: narsil.sondelali.com/docs/ag…
1
29
I'm glad to have tested this as a beta user. It's essentially multiple tools in one. Instead of using Wispr Flow, Notion, and other combined tools, you can just use this.
It’s also more private than the competition, and you pay once and never again. I'm proud of you, @TheSpacerr. Nice job on this one!
🔥 Introducing Koi 👻 🔥
A privacy-focused little ghost that lives on your Mac.
→ An AI that sees your screen, like HeyClicky
→ Dictation that types anywhere, like Wispr Flow
→ Call notes, like Granola
→ Documents, like Notion
→ Flashcards, like Anki
→ One calendar for every account, like Fantastical
One app instead of seven.
Your documents, calls and calendar stay on your Mac.
koiscribe.com
More below 👇
1
5
136
When your coding agent makes the same mistake again and again, will you pay for a stronger model, or write a rule for that mistake?
ALT On a dark square card, the sentence 'Write a rule for the mistake that your coding agent repeats before you pay for a stronger model' appears in large type, with 'Write a rule for the mistake that your coding agent repeats' in blue.
1
4
112
Context7 is a service that puts the latest library documentation into AI agents such as Claude, Codex, and Cursor: context7.com
19
Every customer who uses your AI feature can get an error at once, whenever your spend with your AI provider reaches its monthly limit.
On its Start tier, Anthropic's API has a monthly cap of $500. After that, the API returns an error for every request until the next month begins, and the SDK's automatic retries fail as well.
Before you release:
→ Set a low spend limit on a separate workspace.
→ Use the feature on that workspace until your spend reaches the limit, and look at what your customer will see.
→ Tell your coding agent to show a plain message, and to stop retrying, when Anthropic reports a spend limit, with a 400 error for a limit that you set or a 429 error for the tier's cap. Then repeat the test.
5
7
162
Anthropic documents the spend cap, the error that the API returns once you reach it, and the limits that you can set on a workspace: platform.claude.com/docs/en/…
12
I feel building is so you have something people want to actually buy instead of them paying for a dream
Also I’ve been cold DM’d many times and have never once even replied. If I have a need for something I want to find the perfect thing to fit. So make your product ready and put it out for people to find at their own speed.
1
1
1
28
A customer can be misled to trust a quote from your AI agent because it appears to come with a link to a real page, even when those words are missing from that page.
In a 2026 evaluation of 14 models, for example, the strongest of them scored above 94% on link validity but between 39% and 77% on factual accuracy against the sources in their answers.
To catch that, tell your coding agent to check each quotation against the text that you stored, before your app sends the answer.
The guide in the first comment contains the code for a check that sends an answer only when its quote appears in the stored rule.
4
5
198
A model's answer to a passenger can be 'A folding pushchair travels free', although the rule says 'free of charge for each child under two'. This article teaches how to solve this: narsil.sondelali.com/docs/ag…
28
Do you win more paying customers by messaging potential customers for a weekend than by building with your coding agent for the same weekend?
ALT On a dark square card, the sentence 'Message the people who could pay for your app before you build its next feature, because they can tell you which feature is worth building' appears in large type, with 'Message the people who could pay for your app' in blue.
2
2
65
Sirannon is an open-source tool for hosting your app's SQLite database, with a db.backup call that, where the driver supports it, copies an open database to a file that the tests for each pull request can open as their own copy: sirannon.sondelali.com/docs/…
11