Tldr
TL;DR: Effective Harnesses for Long-Running Agents
Summary of Anthropic's guide on building harnesses for agents that work across multiple context windows.
Long-running agents lose memory between sessions - like engineers working shifts with no handoff notes.
┌─────────────────────────────────────────────────────────────────┐
│ SOLUTION ARCHITECTURE │
│ │
│ Session 1: Initializer Agent │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ • Create init.sh (dev environment setup) │ │
│ │ • Create claude-progress.txt (work log) │ │
│ │ • Initial git commit │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ Session 2+: Coding Agent │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ 1. Read progress.txt + git log │ │
│ │ 2. Work on ONE feature │ │
│ │ 3. Test end-to-end (Puppeteer, not just unit tests) │ │
│ │ 4. Commit with descriptive message │ │
│ │ 5. Update progress.txt │ │
│ │ 6. Leave codebase production-ready │ │
│ └─────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘Key Patterns
- Feature list (JSON): 200+ features to prevent premature "I'm done"
- Incremental progress: One feature at a time, not one-shot everything
- E2E testing: Browser automation catches bugs code review misses
- Session startup checklist: Read dir → progress file → feature list → git log → run tests
Testing with Browser Automation
Open Question
Single general-purpose agent vs. specialized agents (testing, QA, cleanup)?
Reference
- Effective harnesses for long-running agents - Anthropic Engineering