GuideAdvanced

Long-Running Agents: the Harness

An agent that answers a question fits in one session. One that builds a project does not: it works for hours or days, across many sessions, and each session starts without remembering anything about the last one. This guide teaches you to build the **harness**, the system around the model that makes that long work possible: who defines what "done" means, what gets written down between one session and the next, what happens when the process dies halfway, who decides that a feature passes, how much the agent can spend and what it can touch, how the conversation stays inside its context budget, and how you read afterwards what happened. You learn it by building the harness that gets an agent to write **ratecard**, the shipping quote package for Brindle Freight, a fictional parcel company: an initializer agent turns the spec into a feature list with acceptance tests; twelve coding sessions move one feature at a time, with a fixed startup routine, a progress log and a commit at the end; a checkpoint after every step lets you kill the process and resume without sending the same notice twice; a verifier rejects the feature the agent calls done without testing, or that breaks one that already passed; budgets, approvals and the sandbox bound what the agent can do alone; and a trace with a logical clock lets you follow one feature across sessions and replay a whole session without calling the model again. All the code runs on your laptop without an API key, with a deterministic simulated model that has the same response shape as the Claude Messages API and that makes, when you ask it to, the mistakes Anthropic documents in "Effective harnesses for long-running agents": declaring done without testing, touching the test, getting stuck, trying to leave its folder. You see the same numbers as the guide. An optional `--live` mode uses the real API. Every piece is connected to its equivalent in the Claude Agent SDK and the OpenAI Agents SDK, cited from their documentation.