Agent trends: knowledge management and harness engineering

I want to share two hot trends in the world of autonomous AI agents: long-term knowledge management, and setting up full-featured, secure agent environments. This is the next turn of the evolution that makes agents more autonomous and lets us delegate complete tasks to them, even in complex projects.

In February, OpenAI engineers published an article about their internal experiment of building a product 100% with Codex. Along the way they ran into a number of difficulties and went through the hard path of building an effective agent environment. A few key points:

• the environment must have full-fledged observability and quality-control tools;
• project knowledge must be well structured and constantly evolving;
• architectural decisions must be recorded and enforced deterministically.

In his interview on using AI agents, Andrej Karpathy also notes that effective knowledge management and setting up a complete agent environment take the most time when working with AI agents. A few weeks later he published a viral post on X about using Obsidian for knowledge management. Given the huge interest in the topic, Andrej described the idea in detail on GitHub so that anyone can implement it locally with their own AI agent.

And just last week Anthropic launched Managed Agents and published an article rethinking the architecture of the deep agent loop. They clearly separated the agent loop, the agent environment, the tools and the session context. That separation made the environment simpler to set up and more secure. And, of course, Anthropic got a new way to monetise agent infrastructure. They haven’t fully integrated knowledge management yet, but I think that will be solved very soon.

For balance, I’ll finish with a short interview with Michael Bolin about harness engineering and the future of AI agents in software development.