full-agent-in-sandbox
A multi-agent sandbox architecture pattern in which the entire coordinator-worker agent tree runs inside the sandbox, with each sub-agent having an isolated working file system. The agent plane need not be trusted, but cold starts are heavier and observability requires additional plumbing.
Timeline (last 90 days)
The full-agent-in-sandbox pattern is better suited to coordinator-worker structures: each sub-agent has its own working filesystem, and the agent plane itself does not need to be trusted.
The trouble with the full-agent-in-sandbox pattern is heavier cold starts, and getting sub-agent progress out of the sandbox and into LangSmith for observability requires extra pipeline plumbing.
When the agent runs outside the sandbox, LangSmith can automatically capture the latency and parameters of each sandbox round trip; when the entire agent tree runs inside the sandbox, these internal events are invisible to LangSmith, and manual instrumentation or a logging pipeline must be built to get the data out of the sandbox.
Under the sandbox-as-tool pattern, if sub-agent A writes a file and sub-agent B needs to read it in the next step, there are only two options: keep the sandbox warm for the entire sub-agent run (approximately full-in-sandbox), or materialize the state back to the agent filesystem between each call.
In multi-sub-agent scenarios, keeping the sandbox warm effectively blurs the line between sandbox-as-tool and full-agent-in-sandbox, and the engineering complexity is already very close.
All Facts (5)
The full-agent-in-sandbox pattern is better suited to coordinator-worker structures: each sub-agent has its own working filesystem, and the agent plane itself does not need to be trusted.
50%UnverifiedThe trouble with the full-agent-in-sandbox pattern is heavier cold starts, and getting sub-agent progress out of the sandbox and into LangSmith for observability requires extra pipeline plumbing.
50%UnverifiedWhen the agent runs outside the sandbox, LangSmith can automatically capture the latency and parameters of each sandbox round trip; when the entire agent tree runs inside the sandbox, these internal events are invisible to LangSmith, and manual instrumentation or a logging pipeline must be built to get the data out of the sandbox.
50%UnverifiedUnder the sandbox-as-tool pattern, if sub-agent A writes a file and sub-agent B needs to read it in the next step, there are only two options: keep the sandbox warm for the entire sub-agent run (approximately full-in-sandbox), or materialize the state back to the agent filesystem between each call.
50%UnverifiedIn multi-sub-agent scenarios, keeping the sandbox warm effectively blurs the line between sandbox-as-tool and full-agent-in-sandbox, and the engineering complexity is already very close.
50%