---
title: "Comment by Abraam Gotsiridze on \"What Can Your Agent Actually Touch?\" | Hashnode"
description: "The four gates framing is the clearest version of this I've read. \"A page can't do anything by itself, it can only change what the agent decides next\" is the se"
url: "https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d/comment/6aeb4b4159487bd803adb875"
---

[Discussion](https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d)

Comment [What Can Your Agent Actually Touch?](https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d)

[Abraam Gotsiridze](https://hashnode.com/@abraamgotsiridze)

[Nov 4](https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d/comment/6aeb4b4159487bd803adb875)

The four gates framing is the clearest version of this I've read. "A page can't do anything by itself, it can only change what the agent decides next" is the sentence I'll use the next time someone asks if they just need to avoid shady sites.

The row that surprised me most was reads. It's easy to turn on a sandbox and assume your keys are safe, when it's really protecting writes and outbound traffic. Calling out that ~/.ssh is readable by default in every column is worth the whole table.

One habit that pairs well with your AGENTS.md block: split research and changes into separate sessions. Let the agent browse and write up what it found, then start a fresh session from that summary before it edits files or runs commands. The untrusted rule helps, but it still sits in the same context as whatever the page said. A fresh session means the injected text isn't there at all when the risky tool calls happen.

Have you found a clean way to fence in MCP servers that genuinely need network, like a GitHub one, without opening the whole network to them?

[Andre Burgstahler](https://hashnode.com/@r3dtuxedo)

[Nov 4](https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d/comment/6aeb4b4159487bd803adb875#reply-6aeb73fe97b3733a7fde3724)

Glad the reads row landed. The research-then-fresh-session habit is a solid complement to the untrusted-content rule: once the page text is gone from context, it can't steer the edit turn.

On MCP servers that need network: keep the agent host's network deny-by-default, and put only that server in a separate process or container with an allowlist (e.g. api.github.com) and a short-lived, narrowly scoped token. The agent talks to the local MCP endpoint; it never gets a general egress path. If the server must run in-process, expose only the tools you need and treat the token's scopes as the real boundary — same idea as the remote-server row in the table.

[Abraam Gotsiridze](https://hashnode.com/@abraamgotsiridze)

[Nov 6](https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d/comment/6aeb4b4159487bd803adb875#reply-6aedc284fb9b85e68b9f7c23)

That's a clean answer — thanks. Deny-by-default on the agent host, with the networked MCP server alone in its own process or container behind an allowlist and a short-lived scoped token, is the fence I was missing. Treating the token's scopes as the real boundary also keeps a broad allowlist from turning into full access by another name.

[Andre Burgstahler](https://hashnode.com/@r3dtuxedo)

[Nov 6](https://hashnode.com/posts/what-can-your-agent-actually-touch/6ae74292cfccaa26775cc03d/comment/6aeb4b4159487bd803adb875#reply-6aee1cdd5246263530517fcc)

Glad that framing landed. The short-lived scoped token is the piece people skip: once the allowlist is wide, a long-lived broad token is full access with extra steps. If you ever expose more than one networked MCP from the same host, I'd give each its own token (and its own process) so a leak in one server can't reach the others.
