Why long‑running agents matter

• Modern AI agents can write code, test software, and manage projects over many sessions, not just answer one‑off prompts.​
• Without structure, they lose track of what’s been done, redo work, or declare a project “finished” when it isn’t.​
• A good harness lets you trust an agent to make steady, measurable progress toward business goals instead of behaving like an unreliable freelancer.​

The core idea: two agents, one process

Anthropic’s setup uses two roles inside the same system: an initializer and a coding agent.​

• The initializer runs once at the start to set up the project environment, feature list, and tracking files.​
• The coding agent then runs in repeated sessions, making small, safe improvements each time and leaving clear notes and code history for “the next shift.”​

How the initializer sets things up

The initializer’s job is to turn a vague idea like “clone this app” into a concrete, trackable plan.​

• It creates a detailed feature list (often hundreds of specific behaviors) and marks everything as “not done” so the agent cannot claim victory too early.​
• It also creates a progress log, a starter code repository, and an init.sh script so future sessions know how to run and test the project without guessing.​

How the coding agent works safely

Each time the coding agent runs, it behaves more like a disciplined engineer than a “magic box.”​

• It picks one feature at a time, reads the log and git history, updates the code, and leaves the system in a clean, working state before stopping.​
• It documents what changed, commits code with clear messages, and only flips features from “failing” to “passing” after testing them end‑to‑end, often using browser automation tools.​

Key business benefits

For a business owner or exec, this harness design translates into fewer surprises and more predictable results.​

• Work becomes traceable: you can see which features are done, what changed, and what failed, just like with a human dev team.​
• Risk drops because the agent is pushed to test like a real user, keep the app running, and avoid leaving half‑finished, undocumented work behind.​

To dive into the technical details, examples, and failure modes Anthropic studied, read the full engineering post: Effective harnesses for long-running agents