Anthropic's own walkthrough of Claude Managed Agents, explaining the brain/hands split that underpins the platform. It is the best starting point for the four-part…

A hands-on build of a research agent on Claude Managed Agents, led by Anthropic. Useful for seeing the platform work end to end.

Anthropic walks through scheduled agents and the 'dreaming' feature in Claude Managed Agents, with Q&A on running long-lived agents in production.

Anthropic's closing session in the Managed Agents series covers evals, CI/CD and multi-tenancy, the unglamorous parts of shipping agents.

OpenHands' Robert Brennan compares four ways to isolate coding agents and explains why sandboxing is mandatory. A practical talk from a team that ships agents.

Snorkel walks through training a small-model finance agent for under $500, with concrete cost and results numbers.

A talk from Coder on the lethal trifecta (private data, untrusted input, external actions) in agents already running on developer laptops, with practical mitigations.

A full 43-minute freeCodeCamp course on architecting scalable AI automations in n8n from scratch, with real workflows rather than toy demos.

LangChain gives a practical checklist of questions every agent engineer should be able to answer about their system before it ships.

OpenAI's Dominik Kundel walks through building on the Codex harness, the agent loop under Codex that you can extend with your own tools.

Airbyte's engineer explains how to design CLIs that agents can use reliably, with practical interface choices from a team that ships them.

A team explains why they deleted their MCP server and rebuilt it, with practical lessons on tool design for agents.

A LangChain engineer shows how to give agents memory that learns over time, with concrete patterns you can implement.

AWS's Elizabeth Fuentes Leone explains why stuffing more into the context window degrades agents and what to do instead: retrieval, compaction and scoped memory.

Will Lyon of Neo4j shows how to turn an agent's raw memory into reusable skills that actually hold up, with patterns you can copy into your own harness.

LangChain's team explains how they approach systematic agent improvement in Engine v2, a practical look at evals and iteration loops.

OpenAI's short official walkthrough of building plugins for ChatGPT, a quick start on extending it with tools.

Google DeepMind's Philipp Schmid walks through the Gemini Interactions API: server-side state, background jobs and a cloud sandbox for every agent in one call.

Runlayer's founding engineer shows how agents that write their own skills went from 36% to 100% task success, and how to share those skills company-wide.

Nick Saraev breaks down what an AI harness actually is and how operators (not just engineers) can wire one up around Claude Code and similar agents.

Tessl's Dru Knox lays out harness engineering: the control plane, agent-ready stack and improvement loops you need to go from using coding agents to running a software…

Anthropic's own 34-minute session on building secure agents for knowledge work, covering permissions and guardrails you can apply to production agents.

Official LangChain walkthrough of building a model router inside an agent harness, so you can route tasks across models by cost and capability.

Sam Witteveen tests Holo4, an open-weight model that clicks through GUIs, writes and runs code and calls MCP tools, choosing the right action at each step.

LangChain's Interrupt NYC opening keynote lays out where agent engineering is heading and what the team is shipping next across LangGraph and LangSmith.

Browserbase on why LLM-as-judge evals quietly lie to you and how to build verifiers for browser agents you can actually trust.

Laurie Voss (Arize) runs a live improvement loop on a failing shopping agent: add tracing, find the bad tool call, have a coding agent fix it, re-evaluate.

An AWS engineer walks through agents that generate and register their own tools on the fly, with the guardrails needed to make that safe in production.

OpenAI's head of ChatGPT and Codex, Tibo Sottiaux, on Dots, why most internet actions will soon be taken by agents, and why hand-built agent graphs are a passing phase.

A full code-along building an agentic harness in Python from scratch: the loop, tool calls, state and guardrails, with no framework hiding the mechanics.

Hugging Face rebuilds OpenAI's always-on dots agent with open models, Pi and a Telegram front end, so you can run your own personal agent end to end.

OpenAI's full live demo of dots, the always-on agents announced at DevDay 2026, showing how they pick up tasks, use tools and run unattended.

Latent Space sits down with OpenAI's Ari Weinstein and Nikunj Handa to unpack the post-DevDay agent stack: computer use, the Decisions API, UltraFast and always-on dots.

Fireship explains the capability-based security idea borrowed from a 1970s military design and why it may be the most practical defense against agent prompt injection…

Hands-on code walkthrough of wiring Jev, TypeSafe AI's decision-only model that returns typed choices instead of text, into an agent harness for cheap, fast routing.

a16z partners put the new generation of personal AI agents through real tasks and break down where they hold up, where they fall over, and what that means for builders.

End-to-end walkthrough of the Claude Agent SDK: tools, sessions, permissions and shipping an agent that does real work. Start here if you have used Claude Code and want…

Five minutes on the three primitives that make Claude Code scale past one terminal: skills, subagents and handoffs. Watch before you set up a team's shared Claude Code…

A multi-agent system in n8n: an orchestrator that hands off to specialist agents, with memory and error handling. Good pattern for internal automations before you write…

n8n's official quick start: wire an LLM to tools and memory in a visual workflow and deploy it. The fastest way to put an agent behind a Slack channel or a webhook…

Anthropic's own workshop on the Agent SDK: the agent loop, tool design, subagents, context management and evaluation. Nearly two hours, from the person who built it.