How Hightouch built their long-running agent harness (15 minute read)
Hightouch developed an AI marketing agent capable of complex, multi-step tasks. Its team explicitly separated the model's planning from its execution, allowing for dynamic plan updates as the agent learns from data. To manage context, they used "agentic delegation" through techniques like buffering large data to a temporary filesystem and spawning dynamic subagents for isolated, complex sub-tasks. |
Unconventional PostgreSQL Optimizations (20 minute read)
This article goes over some unconventional PostgreSQL optimization techniques that go beyond standard practices like simple indexing. One method involves enabling `constraint_exclusion` to remove unnecessary full table scans when queries contradict existing check constraints. Another technique optimizes for lower cardinality by creating function-based indexes (e.g., on just the date part of a timestamp), which are made consistently usable and efficient with virtual generated columns. |
|
I'm addicted to being useful (5 minute read)
A staff software engineer deeply loves his job despite current industry stresses because of his addiction to being useful. He gets satisfaction in work that aligns with his compulsion to solve problems. His role is perfectly tailored to this, as he feels an almost physical discomfort until he resolves a technical challenge. |
AI's Missing Ingredient (6 minute read)
The current reliance on basic chatbot interfaces underutilizes AI's potential. This failure of imagination comes from the absence of design. AI needs more design expertise to create better products. |
|
Announcing Mastra 1.0! (7 minute read)
Mastra, an AI agent framework for TypeScript, has officially announced the stable release of version 1.0. This version includes Server Adapters for easier deployment, Composite Storage for flexible backend configurations, and full support for Vercel's AI SDK v6. |
Pipenet (Website)
Pipenet is a modern, open-source alternative to localtunnel that provides a bundled client and server to host your own secure tunnel infrastructure. It allows users to expose local services to the internet, share local development servers, test webhooks, or embed tunneling capabilities directly into their own applications. |
wxpath (GitHub Repo)
wxpath is a Python library for declarative web crawling and data extraction, where traversal and extraction logic are expressed directly using XPath. It introduces custom XPath operators like `url(...)` and `///url(...)` to enable deep, breadth-first web graph exploration within a single expression. |
|
Without Benchmarking LLMs, You're Likely Overpaying 5-10x (9 minute read)
Many users overpay for LLM API calls because common benchmarks don't accurately predict performance or cost-effectiveness for their specific tasks, leading them to default to popular but expensive models. This article goes over a benchmarking methodology that involves collecting real prompts, defining expected outputs, running numerous LLMs via a unified API, and using an LLM-as-judge to score responses based on quality, cost, and latency. This process consistently revealed that many users can achieve comparable quality at much lower costs, often cutting bills by 80% or more. |
AI Agents at work: real-time platform insights in Slack (8 minute read)
Monday.com developed an AI Slack bot to help devs quickly access real-time platform insights without having to manually check internal tools or dig through notifications. The bot uses LLM agents connected to their microservices data through an MCP. To prevent the AI from hallucinating data, they pre-computed statistics rather than letting the model calculate them. |
Running Claude Code dangerously (safely) (6 minute read)
The author wanted to use Claude Code with the `--dangerously-skip-permissions` flag for an improved workflow, but needed a safe, isolated environment to prevent accidental damage to their host system. They initially considered Docker, but issues like Docker-in-Docker requiring privileged mode made it unsuitable for true sandboxing. They chose Vagrant to create a disposable virtual machine, which offers full isolation and a reproducible configuration. |
|
Claude Chill (GitHub Repo)
Claude Chill is a PTY proxy that tames Claude Code's massive terminal updates by rendering only screen differences, preserving history, and enabling lookback to prevent lag, flicker, and broken scrollback. |
The Agent Skills Directory (Website)
Skills are reusable, plugin-like capabilities that provide procedural knowledge to help AI agents more effectively accomplish specific tasks, and this website is a directory of the most useful skills. |
|
| Love TLDR? Tell your friends and get rewards! | | Share your referral link below with friends to get free TLDR swag! | | | |
| Track your referrals here. |
|
| |
0 Comments