Sheva · Part 4 · AI Integrations

Run It Safely.
Agents in the Real World.

Letting agents do more — without letting them do damage.

Two tracks · Everyday for everyone, Developer for the builders.
CORE IDEA

Limit the blast radius

EVERYDAY TRACK

Safe guardrails, no code

DEVELOPER TRACK

Sandboxes & test environments

Pick Your Track

This one splits in two.

Agents that act on your behalf are powerful — and that power needs boundaries. The principle is the same for everyone: give an agent room to help, but wall off what it could hurt. How you do it depends on who you are.

Tap a track to jump to it — or just scroll; both are below.

Everyday Track
E1 · The Big Idea

Limit the blast radius.

When an agent can open files, browse the web, and run tasks for you, the goal is simple: if it ever makes a mistake — or gets tricked — the damage stays small and contained.

Give it a sandbox

A dedicated folder for agent work — copy files in, don't point it at your originals.

Give it a short leash

Only the access it needs for the task. Not your whole drive, inbox, or accounts.

Watch it first

Supervise the first few runs before you trust it to work unattended.

Know the off switch

Be able to pause it, revoke access, and stop a scheduled task at any time.

Everyday Track
E2 · The Practical Rules

Do this. Not that.

These are the habits that prevent almost every everyday agent mishap.

Do
  • Make one dedicated "agent work" folder and copy files into it.
  • Tell it plainly what it may not touch: "don't delete anything."
  • Be specific about the outcome — vague tasks invite guesswork.
  • Limit any browser helper to sites you trust.
  • Review its output before anything gets sent or published.
Don't
  • Put passwords, financial files, or credentials in the working folder.
  • Give it standing access to your whole drive or inbox "just in case."
  • Let it act on untrusted web pages, emails, or documents unchecked.
  • Assume org settings protect you on a personal account — they may not.
  • Trust an unattended, scheduled agent you've never watched run.
Why the caution?
Agents that read the web and your files can be fooled by hidden instructions buried in a page or document — "prompt injection." It's the #1 risk. The folder trick and a short leash are what keep a bad instruction from becoming a bad outcome.
Everyday Track
E3 · At Work

If it's for your job, loop in IT.

Workplace agents touch company data — so the rules aren't just yours to set.

1
Use only approved tools and connections

Your organization vets agents for a reason. Don't wire one into company systems on your own.

2
Assume someone should be able to audit it

Know what your agent did and when. If you can't show your work, slow down.

3
Ask before you automate anything sensitive

Customer data, finances, personnel info — get a yes from the people who own the risk.

The everyday takeaway:
You don't need to be technical to be responsible. A dedicated folder, a short leash, a quick review, and looping in IT at work — that's 90% of safe agent use.
Developer Track
D1 · Test Before You Trust

Never let an agent run loose on your real machine.

Coding agents write and run code you haven't reviewed. The pattern in 2026 is to run them in an isolated environment — a sandbox — so a bad command destroys a disposable container, not your system.

Isolation

Own filesystem, network, and process space — walled off from your host, credentials, and production data.

Ephemeral

Spin up, run, tear down. Nothing persists unless you choose it to.

Reproducible

Same environment every run — so a test result actually means something.

Recoverable

If it breaks, you throw it away and start fresh. No cleanup on your real box.

Developer Track
D2 · The Tooling Landscape

Where people run agents safely.

A fast-moving space — verify current specs before you commit. As of mid-2026, roughly:

Agent-native sandboxes (built into the coding agent)

Isolate for safety while the agent writes and runs code. Great for generation; the "last mile" of real deployment/verification often stays manual.

Claude CodeOpenAI CodexCursor (cloud agents)GitHub Copilot agent

Dedicated sandbox infrastructure (bring your own agent)

Purpose-built isolated compute — microVM-level separation, SDKs, fast boot. For running untrusted code at scale.

E2BModalDaytona (open-source)Fly.io SpritesVercel SandboxCodeSandbox

Self-hosted dev environments (your own hardware)

Run environments on your own servers — or that old machine in the closet — for full control and data residency.

CoderDaytona (self-hosted)Docker / dev containersNorthflank (BYOC)
Sandboxes solve execution — not verification.
Most can run tests, but can't hit real databases or hand you a working preview. That last mile — proving it works end-to-end — still needs a human. Isolation buys safety, not certainty.
Developer Track
D3 · Guardrails in Code

Permissions ask. Hooks enforce.

For production agents, "please don't" isn't a control. Two layers matter:

1
Permissions / allow-lists

Define a narrow set of tools and actions the agent may use. The model still decides within them.

2
Deterministic hooks

Your own code firing at fixed points — block destructive commands, inject context, audit every action. The final word, not a suggestion.

3
An enforcement point outside the model

A proxy or gateway that intercepts agent traffic. Because you can't trust the thing you're trying to constrain to constrain itself.

The developer takeaway:
Isolate execution, constrain with allow-lists, enforce with code the model can't override, and keep a human on the last mile. Same "limit the blast radius" idea as the Everyday track — just with real teeth.
Both Tracks · The Through-Line

Powerful and careful aren't opposites.

Whether you're copying files into a folder or wiring up a microVM, it's the same instinct: let the agent help freely inside a space where a mistake can't hurt you.

The one thing to remember:
Give agents room to work and walls around the damage. That's the whole discipline — at every level.

This is the frontier, and it moves fast — treat specific tools as a snapshot, not gospel. The principle is what lasts.

The AI Series · Your Path
1 Choose a Tool 2 Now What? 3 Now Build It 4 Run It Safely