Effective Harnesses for Long-Running Agents
Source: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents (inferred) Author: Anthropic Engineering Date: 2025-11-28
Summary
Anthropic’s engineering guide on building harnesses for long-running agents. Companion to the Harness Engineering post from OpenAI/Fowler. Anthropic’s framing: harnesses are not just scaffolding — they’re where most of the engineering value lives.
Key Claims
- Harness responsibilities: checkpointing, error recovery, context management, tool routing, logging.
- Long-running means: tasks that outlast a single context window — the harness must manage context refreshes.
- Checkpointing pattern: save agent state at key decision points so failures can resume, not restart.
- Error taxonomy: (1) model errors (hallucinations, misunderstandings), (2) tool errors (API failures), (3) harness errors (infra). Each needs different recovery.
- Key principle: the harness should be deterministic even when the model isn’t — reproducibility lives in infrastructure.
- Context refresh strategy: summarize old context rather than truncate — preserve decisions and rationale.
Entities
Concepts
- Coding Agents — long-running coding agents are the primary target
- Context Engineering — context management is core harness responsibility
- Agent Memory — checkpointing is a form of episodic memory