Anthropic Sandbox Runtime adapter

Status: Experimental. AgentDiff can wrap argv with a separately installed srt executable. AgentDiff records that external sandboxing was requested; effective enforcement depends on the installed runtime, operating system, version, and settings.

The adapter does not bundle srt, translate AgentDiff policy into sandbox policy, parse enforcement logs, or prove which controls were applied.

CLI

Install and configure srt separately, then run:

agentdiff run \
  --runtime srt \
  --srt-executable srt \
  --srt-settings /absolute/path/to/settings.json \
  --task "Fix the parser" \
  -- python agent_task.py

The executed wrapper argv is:

srt --settings /absolute/path/to/settings.json -- python agent_task.py

Use the normal --timeout option for the overall observed command. There is no --sandbox-timeout or --sandbox-logs flag.

Python

import sys

from agentdiff import AgentRunTransaction, load_policy_file
from agentdiff.runtime import SandboxRuntime

runtime = SandboxRuntime(
    root="/workspace/project",
    executable="srt",
    settings="/absolute/path/to/settings.json",
)
result = AgentRunTransaction(
    root="/workspace/project",
    policy=load_policy_file("/workspace/project/agentdiff.yaml"),
    task="Fix the parser",
    runtime=runtime,
).run([sys.executable, "agent_task.py"], timeout_seconds=300)

print(result.runtime.backend)      # anthropic-sandbox-runtime
print(result.runtime.enforcement)  # external_sandbox_requested

Boundary

AgentDiff captures host-visible before/after state around the wrapper. It does not claim that host-visible observations equal all in-sandbox activity. Review the external runtime's own settings and guarantees independently.