Agent SafetyOutlineDraft

Building the Sandbox: Isolated Tool Execution

Unsafe tools should never run on the agent host. Dispatch them to ephemeral throwaway sidecars and destroy the workspace after.

Series draft — Part 7 of 15 in Hardened Agentic Stack. Outline only; expand before un-drafting.

Phase 2: Runtime Integrity — The Local Threat Model

The Problem

Direct filesystem/network access for tools is inherently unsafe on a developer or edge host.

The Infrastructure Fix

Route unsafe tools through ephemeral container sidecars (Kata/gVisor where needed).

The Architecture Pattern

Ephemeral Execution Sidecar — agent dispatches; container dies after the task.

Planned sections

  1. The “Oh No” moment — concrete incident or near-miss that makes the risk visceral.
  2. ClawQL context — how this control protects a high-privilege local/edge agent.
  3. Technical how-to — concrete configs, policies, or snippets a builder can apply.
  4. Safety check — what “trusted enough” looks like once this layer is in place.

Key visuals

  • Agent → sidecar → destroy lifecycle

Source modules (docs.clawql.com)

Rule of Three (keep on publish)

LayerTakeaway
ProblemDirect filesystem/network access for tools is inherently unsafe on a developer or edge host.
Infrastructure fixRoute unsafe tools through ephemeral container sidecars (Kata/gVisor where needed).
Architecture patternEphemeral Execution Sidecar — agent dispatches; container dies after the task.

About the author

Daniel Smith builds ClawQL, an agent operating system for token-efficient discovery and execution over APIs — with observability, hardened tool boundaries, and production routing for LLM workloads. He writes here about the systems problems behind shipping agents.