Break Your Agent with Chaos Testing, Then Fix the Harness

Resilient AI Agents: A Harness That Catches What the Model Can't

Open-source repo: five runnable demos (chaos testing, memory guardrails, poisoning defense, multi-step planning, self-improving skills) plus the AgentCore runtime

If you find it useful, give it a ⭐!

Presentation Deck

Full slide deck: the four production failures and the harness control each one needs

Stop AI Agent Hallucinations: Validate Before the Agent Writes to Memory

The before-write memory guardrail in depth

How to Stop AI Agents from Hallucinating Silently with Multi-Agent Validation

Executor, validator, and critic: catch silent failures before the user does

AI Agent Guardrails: Rules That LLMs Cannot Bypass

Enforce rules in code at the tool boundary, not in the prompt

PALADIN: Self-Correcting Language Model Agents to Cure Tool-Failure Cases

Research behind the chaos-testing diagnose-fix-validate loop

Zombie Agents: Persistent Control of Self-Evolving LLM Agents

Research on memory poisoning as a persistent, cross-session compromise

MiRA: A Subgoal-driven Framework for Improving Long-Horizon LLM Agents

Research behind the execute-then-verify loop for multi-step tasks

Memento-Skills: Let Agents Design Agents

Research behind the self-improving, write-test-reuse skill loop