Harness logic that compensates for model weaknesses goes stale as models improve. Separating session log, harness loop and sandbox behind stable interfaces lets you swap each one without rewriting the others.
Key takeaways · AI-distilled
Anthropic first ran session, agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → and sandboxAn isolated environment where AI-generated code or agent actions run without being able to touch anything real.Full definition → in one container, a 'pet': a harness bug, a dropped stream packet and a dead container all looked the same, and debugging meant opening a shell beside user data. Now the harness calls the container as just another tool.
Because the session log lives outside the harness, Anthropic treats harnesses as disposable: after a crash a new one boots with the session ID, fetches the event log and resumes from the last event. A dead container returns a tool error, and Claude can retry on a fresh one.
Anthropic keeps credentials out of the sandbox where Claude's code runs instead of relying on narrowly scoped tokens. Git tokens are wired into the repo remote at setup, and MCP OAuth tokens sit in a vault behind a proxy, so neither Claude nor the harness handles them.
Anthropic splits durable context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → storage from context management: the session log keeps every event and serves positional slices for rewinding or rereading, while context compactionSummarizing an agent's earlier conversation to free room in the context window so a long session can keep going.Full definition → and trimming happen in the harness, so irreversible pruning never loses the record.
Provisioning containers only when Claude makes a tool call, instead of booting one per session before inference, cut Anthropic's median time-to-first-token by roughly 60% and its p95 by over 90%, since sessions that never touch a sandbox no longer pay setup cost.
Terms in this piece · Glossary
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
