## Definition **Subagent context isolation** is the architectural property that each subagent in the [[Orchestrator-Subagent Pattern]] runs with its own fresh, independent context window. It receives only the task description from the orchestrator, executes in private, and exposes nothing of its internal state back except the [[Compressed Summary Return]]. It can't read the orchestrator's history, peer subagents' state, or other parts of the run. ## Why Isolate Three concrete failure modes the isolation defends against: - **Context bleed**. Without isolation, every subagent inherits an ever-growing transcript. By the third or fourth spawn, the cheapest model has a 200k-token prompt full of irrelevant tool-call history. Cost and latency degrade linearly with the depth of the run. - **Cross-task contamination**. A subagent investigating topic A picking up partial reasoning about topic B will conflate them. Empirically this is one of the most common quality regressions in early multi-agent systems. - **Failure containment**. If a subagent corrupts its context (prompt injection from a fetched page, runaway loop, hallucinated assumption), isolation keeps the damage local. The orchestrator gets one bad summary, which is easier to detect and discard than a poisoned shared transcript. ## What "Isolated" Means Concretely In production deployments the isolation is typically literal: subagents run as separate API calls, sometimes separate processes, occasionally separate VMs (Cognition's Managed Devins each get their own virtual machine). The orchestrator passes a fresh system prompt and a focused task description; everything else is empty. No shared memory, no shared file handles, no peer channels. Only the result string crosses the boundary. ## Limits Isolation costs information. A subagent can't ask "what did the user want originally?" without the orchestrator re-supplying that context in the task description. Two subagents can't notice they're duplicating work. The orchestrator becomes the sole point of cross-cutting awareness, and it has to be designed for that. For tasks where subagents genuinely need to coordinate (long-running negotiation, distributed simulation), the isolation pattern breaks down and the GroupChat alternative resurfaces — at the cost of all the problems isolation was meant to solve. ## Related - [[Orchestrator-Subagent Pattern]] - [[Compressed Summary Return]] - [[Ephemeral Subagent]] - [[Prompt Injection]] - [[Context Window]] ## Sources - [[Multi-Agent AI Systems in 2026 (FlowHunt)]] - [[The Architecture of Scale - Anthropic Sub-Agents (Oswal)]]