AI agents
How agents are built, tuned and run: loops, harnesses, MCP, skills and coding agents, with what broke and what we measured.
Best notes
- Agentic engineering: the series map and how to read it
A series on building agentic software for engineers who know the classic lifecycle, and how its pillars, deep dives and reading paths fit together.
- Owner agents, plain functions: where the model sits in our ops estate
Give each domain one owner agent that reports back, keep the steps as plain functions, and add autonomy to a running platform one gated decision at a time.
- 3,016 delegations to agents: the long briefs cost the most
Median brief 289 words, median run 6.4 minutes, 14.5% needed a follow-up. What 24 days of Claude Code subagent logs say about briefing agents, and where a brief cost…
- Harness engineering notes from a team running agents all day
Nine field notes on running coding agents in production at Dwarves: the harness around the model, how we check and guard it, what we put it to work on, and what it does…
- Five layers to tune before you fork your agent
What makes AI agentic, how to split a deterministic core from an agentic edge, and the five layers you can tune without rebuilding the harness. Real numbers from circle…
Reports
- Making our site AI-agent-ready, 50 to 93
We ran memo.d.foundation through is-agentic.com, read its rubric closely, and shipped the fixes that took the score from 50 to 93. Here is the rubric in full and what…
- AI knowledge automation
Exploring the motivations behind the MCP Playbook, its role in establishing a dynamic runbook, and new patterns for capturing AI-generated knowledge like chat logs and…
- MCP playbook data flow
How data flows through the MCP playbook server, from MCP client requests to tool execution and interactions with the file system and GitHub. This overview helps you…
- Building chatbot agent to streamline project management
A technical case study detailing the implementation of an AI chatbot agent in a project management platform. Learn how the team leveraged LangChain, LangGraph, and GPT-4…
- Building data pipeline for OGIF transcriber
A technical case study of creating an automated system that downloads videos, processes audio, and generates transcripts using AI services like Groq and OpenAI.
- Project reports system: a case study
An in-depth look at Dwarves' monthly Project Reports system - a lean, efficient system that transforms communication data into actionable intelligence for Operations…
Writing
- Agentic engineering: the series map and how to read it
A series on building agentic software for engineers who know the classic lifecycle, and how its pillars, deep dives and reading paths fit together.
- Owner agents, plain functions: where the model sits in our ops estate
Give each domain one owner agent that reports back, keep the steps as plain functions, and add autonomy to a running platform one gated decision at a time.
- No. 43: Tuan Dao on delegating to AI
Tuan Dao, an engineer at Dwarves, on moving from pair programming with AI to delegating to it, and why human review is always the bottleneck.
- 3,016 delegations to agents: the long briefs cost the most
Median brief 289 words, median run 6.4 minutes, 14.5% needed a follow-up. What 24 days of Claude Code subagent logs say about briefing agents, and where a brief cost…
- Harness engineering notes from a team running agents all day
Nine field notes on running coding agents in production at Dwarves: the harness around the model, how we check and guard it, what we put it to work on, and what it does…
- Five layers to tune before you fork your agent
What makes AI agentic, how to split a deterministic core from an agentic edge, and the five layers you can tune without rebuilding the harness. Real numbers from circle…
- Dispatch #2: handoff pattern: từ ảnh chụp app đến Figma và video demo
Dùng pattern prompt end-to-end test để AI tự chụp ảnh app rồi convert sang Figma, thay vì test tay. MCP Bridge Figma UI của một bạn Việt Nam trên GitHub giúp tiết kiệm…
- What the Engelberg report got right about us
Thoughtworks published the findings of its second Future of Software Engineering retreat. Most of what the room converged on is what we have been learning and putting…
- Claude Code 2.1.259 turned our Read() deny rules into a Bash tripwire
Bypass-mode sessions started stalling on 'a Read() deny rule is configured' prompts. What the harness now infers from a Bash command, why a CLAUDE.md rule couldn't stop…
- The debug loop: make the agent measure the browser before it touches the CSS
A repeatable loop for AI-driven UI debugging: name the metric, name the axis, drive a real browser over CDP, read the DOM into a table, kill theories, sweep the suspect,…
Show all 25 pieces
- DeepSeek harness, dissected
An architecture read of dsh, DeepSeek's new agent harness: the plugin tree, the patch-layer boot, the log-first session model, and what it means for a Claude Code shop.
- Stagehand breakdown
A comprehensive technical breakdown of Stagehand, an advanced browser automation framework by Browserbase
- Context7 breakdown
Technical analysis of Context7, an intelligent documentation indexing and retrieval system that transforms raw library docs into AI-optimized, ranked snippets for…
- E2b breakdown
Technical analysis of E2B, a cloud infrastructure platform that runs AI-generated code in secure, isolated sandboxes using lightweight virtual machines that start in…
- Dify breakdown
Technical analysis of the Dify platform, its architecture, and engineering decisions that enable scalable LLM application development.
- Mem0 & Mem0-Graph breakdown
Technical analysis of Mem0, a scalable memory architecture for LLMs, and its graph-based variant, Mem0-Graph, designed for long-term conversational coherence.
- Cline breakdown
Comprehensive technical analysis of Cline's VS Code extension architecture, covering system design, implementation patterns, and architectural innovations
- Ax framework breakdown
Technical analysis of the Ax TypeScript framework for building LLM-powered agents with DSPy capabilities.
- Crawl4AI breakdown
Deep dive into Crawl4AI's architecture, data structures, and algorithms - from async pipelines and strategy patterns to browser management and intelligent content…
- Zen MCP breakdown
Technical analysis of the Zen MCP (Model Context Protocol) Server architecture, implementation, and design patterns.
- Composing the future, part 1: agent basics, LangGraph and Autogen
What makes a system agentic, why multi-agent beats single-agent, how orchestration works, and a close look at LangGraph and Microsoft Autogen.
- Composing the future, part 2: OpenAI's agent tools, MCP and fast-agent
How OpenAI's Swarm and newer agent tools work, why Model Context Protocol matters for tool access, and how fast-agent builds MCP-native agents.
- Composing the future, part 3: Mastra and CrewAI
Two frameworks with opposite instincts: Mastra for TypeScript teams building web-native agents, and CrewAI for role-playing crews of Python agents.
- Composing the future, part 4: LlamaIndex and the framework face-off
LlamaIndex's agent features for document-heavy work, then a side-by-side comparison of the frameworks on differentiators, strengths and ease of development.
- Composing the future, part 5: agent patterns, open problems and how to choose
The common agent composition patterns, the problems that remain unsolved, where the field is heading, and guidance for picking a framework.
Research notes
- The Coding Agent Team
I've been experimenting with making AI assistants work not as a single tool, but as a specialized team. It seems to work surprisingly well.
- How we crafted the OGIF summarizer bot to streamline weekly knowledge-sharing
Introducing the OGIF Memo Summarizer, a chatbot we developed using Dify in collaboration with @innno_. This tool transcribes YouTube videos and extracts key points from…
- How we created an AI powered interview system using OpenAI's ChatGPT
Discover InterviewGPT, an AI-powered ChatGPT system that automates technical interviews by generating tailored questions and evaluating answers for engineers at all…
- How knowledge work organizes itself
How societies naturally organize around different types of knowledge work; from those who apply compressed rules to those who derive new principles from scratch.
- The five stages of learning
How human cognition and artificial intelligence both develop increasingly abstract representations, from simple stimulus-response patterns to deep hierarchical…
- Should AI agents transact without human confirmation?
A UX guide to designing agentic AI systems that balance automation with user control.
- AI agent explained: LLMs, workflows and agents
Most people confuse LLMs, workflows, and agents, but understanding these three levels unlocks the real potential of AI automation. The key difference lies in who makes…
- Designing software for AI agents
As AI agents become inevitable, software design must shift from human-centric to agent-centric experiences. The future belongs to systems that optimize for…
Show all 20 notes
- Vibe culture
AI tools handle technical complexity, freeing creators to focus on emotional resonance and creative flow. We explore how this delegation shapes modern creative processes.
- Mapping the path to AGI
Dissecting the evolution of AI agents from eager human-in-the-loop systems to fully autonomous entities, and how each stage automates the core tasks of exploration,…
- Authorization for remote MCP servers over SSE
This guide explores implementing robust authorization for Model Context Protocol (MCP) over Server-Sent Events (SSE) transport, providing a standardized framework for…
- Tool-level security for remote MCP servers, part 1: designing role-based access control
Why remote MCP servers need tool-level access control, the security architecture, and the data model and tool registry behind role-based access.
- Tool-level security for remote MCP servers, part 2: enforcement, identity and practice
How an MCP server enforces permissions and filters data, connects to existing identity systems, and rolls out access control in practice.
- Human touch in agent systems
As agents become capable of making most decisions independently, understanding when and how humans should intervene becomes critical for building effective agent-human…
- Evaluate chatbot agent by user simulation
When building a chatbot agent, it's important to evaluate its performance and user satisfaction. One effective method is user simulation, which involves creating virtual…
- Building agent supervisors to generate insights
In the rapidly evolving field of artificial intelligence, the concept of agent supervisors has emerged as a powerful approach to orchestrating multiple AI agents for…
- Build custom AI agent with ElizaOS
This guide shows how to build a custom AI Agent with ElizaOS.
- Multi-agent collaboration for task completion
In AI integrated systems, instead of put all workload on a single agent, we can apply divide and conquer strategy to distribute workload to multiple agents. This…
- Function calling in AI agents
Function calling is a critical component in the architecture of AI agents, facilitating the integration of external functionalities and resources. This note explores how…
- Explaining gradient descent in machine learning with a simple analogy
Gradient descent is a fundamental optimization algorithm in machine learning. It's a way for models to learn from data and improve their accuracy by gradually adjusting…
Playbooks
- MCP playbook setup
A step-by-step tutorial on configuring and running the mcp-playbook server, including GitHub token creation and using npx. This guide helps you get started with…
Series
Tags in this topic
Writers
monotykamary17, tieubao14, minhluuquang3, nnhuyhoang3, zlatanpham3, chinhld122, hthai22012, innnotruong2