Agent harness is a software layer surrounding a model that manages context, tools, state, and task execution. It determines how agents handle timeouts, process interruptions, and maintain consistency during task execution.
Codex, Claude Agent SDK, Gemini CLI, and Hermes Agent provide pre-built agent environments. LangGraph and Google ADK 2.0 focus on process orchestration through code, while Harness² acts as an experimental layer for modifying the harness itself.
Gemini CLI features a checkpointing system that saves project snapshots and conversation history, allowing users to restore state via the /restore command. Hermes Agent utilizes a skill system to store procedures, with an optional write_approval feature for human review of changes.
LangGraph separates state persistence into a checkpointer for specific threads and a store for cross-thread data. Google ADK 2.0 uses workflows to separate execution routing from language processing, allowing developers to integrate LLM agents with deterministic functions.
Harness² is an experimental project from Google Cloud AI Research that uses a Propose–Probe–Compose cycle to refine agent harnesses. It allows for the modification of eight components, including system prompts, memory, and skills, based on performance in benchmarks like WorkBuddy-Bench.
The effectiveness of these tools depends on the implementation of robust contracts, such as idempotency keys for external operations. Developers must ensure that authorization and validation logic reside in trusted code rather than relying solely on model-generated decisions.
