Context isolation gives a sub-agent a fresh context window for a narrow task instead of passing along the parent’s conversation and accumulated tool results. The sub-agent works in the shared environment and returns its findings, keeping task-specific output out of the main conversation. section 3.1.3 section 3.4.3
Isolation is not the same as an empty environment. In Claude Code, project memory still loads for sub-agents, although the built-in Explore and Plan agents skip CLAUDE.md for speed. The boundary separates conversations, not necessarily all instructions or environment access. section 3.1.3
Route outputs according to their estimated signal-to-noise ratio: the share useful to the next reasoning step.
- Return high-signal outputs, such as git status and targeted file reads, directly. section 3.1.3
- Consider isolation for mixed-signal output, such as repeated build errors or lint runs, when it would crowd the main context. section 3.1.3
- Isolate low-signal streams, such as broad test suites and log analysis, and return a structured summary. section 3.1.3
Treat these bands as operational labels, not fixed thresholds. Sample representative outputs, label useful spans, and calibrate routing against task success, latency, and token cost. A disposable context isolates one task; a persistent index instead amortizes discovery across recurring work. The choice depends on recurrence, codebase size, freshness, and the agent’s state. section 3.1.3 section 3.4.3
Isolating bug-fix output
In an observed bug-fix task, disposable sub-agent windows absorbed search and test output while the main conversation stayed substantially smaller. That is workload-specific evidence, not a guaranteed saving; validate the effect using task-success, latency, and token traces. section 3.1.3

