Self-Improving Agents

It's expensive and time consuming to change, tune, and retrain an LLM to modify an agent’s behavior. Self-Improving Agents: How to engineer adaptive agent harnesses shows you how to design AI agents that measurably get better in production without fine-tuning, weight updates, or waiting for the next model release. Written by Micheal Lanham, author of AI Agents in Action, this hands-on book equips you with techniques to transform your agent’s harness--context, memory, metacognition, tools, and code-- into a measured, versioned, auditable improvement loop that adapts and improves as it runs.

As you go, you’ll build HelixAgent, a RAG general-knowledge agent that starts static and gains a new self-improvement layer in every chapter. By the final chapter, your agent will ship with drift detection, rollout controls, and a reward-hacking runbook. Along the way, you’ll build a portfolio of self-improving agents: a data-analyst agent scored by exact ground truth, a helpdesk agent with four-tier memory that learns from its own traffic, a Karpathy-style hill-climbing research agent, and coding agents whose skill files, tool descriptions, and planner code become the artifact under search.

Throughout, the Helix Observatory web dashboard lets you replay lineage trees, diff candidate contexts, inspect judge verdicts, and stream a live search as it runs. You’ll gradually work your way up to HyperAgents—a cutting-edge research pattern that improves the agent and also the harness self-modification process. By the time you’re done, you’ll have agents running under one improvement loop, applied layer by layer up the agent harness, held to production standards of auditability, human gates, and honest costs.