Agent harness: the 5 pieces your AI agent is missing
Notes de l’épisode
An agent harness is everything around the model that turns it into a working AI agent, and most agents are missing key pieces of it. In this episode we go through five of them: context management, sandboxing, fine-grained policies, fast decision models, and a harness you can run in production. Romain sits down with Morgan Willis, Senior Developer Advocate at AWS, and Arron Bailiss, Principal Engineer and Tech Lead for Strands Agents, to look at what it takes to build a state-of-the-art agentic system today, and how Strands Agents grew from an agent loop SDK into a full toolkit for it. The 5 pieces: 1. A harness with built-in tools, memory and a CLI, open source as Strands harness or managed as AgentCore harness 2. Context management that offloads tool results and compacts history, for 28% lower token cost on the same model 3. A sandbox that limits what files and network your agent can reach (Strands Box) 4. Policies that remember what the agent already did, so it can't issue a refund before checking eligibility (Dogwood) 5. Small decision models that pick between options in milliseconds instead of a full LLM call (Strands Decider) Plus: real-time voice agents with bidirectional streaming, and why Morgan turned a Big Mouth Billy Bass into one. Chapters: 00:00 Strands one year after going open source 02:45 Why model-driven agents get better with every model 13:50 What is an agent harness? Model plus harness 18:20 Piece 1: a harness with tools, memory and a CLI 25:50 Piece 2: context management and 28% lower token cost 29:45 Piece 3: sandboxing agents with Strands Box 31:00 Piece 4: Dogwood policies that track what the agent did 42:45 Piece 5: Strands Decider and small decision models 51:15 Bonus: voice agents with bidirectional streaming 56:45 What's next: agent safety and self-improvementWith Morgan Willis, Senior Developer Advocate at AWS ; With Arron Bailiss, Principal Engineer at AWS and Tech Lead for Strands Agents