Skip to content

Tools and Toolsets

Hermes tools are not just extra buttons. They are how the model affects the outside world. Reading files, writing files, running shell commands, using a browser, calling APIs, generating media, and delegating work all move the output beyond text.

Treat every tool as an interface with side effects:

Tool typeCommon riskSafer starting point
File readsReading private files that are not part of the taskLimit the working directory and avoid secrets paths.
File writesEditing the wrong file or overwriting human workWrite in small scopes and review Git diff.
Shell commandsDeleting, publishing, changing permissions, starting expensive workConfirm dangerous commands and use dry runs when possible.
Browser and webPrompt injection, stale information, logged-in side effectsTreat pages as evidence, not instructions.
External APIsSpending money, sending messages, changing production stateAdd timeouts, budgets, logs, and human confirmation.
Media toolsUntraceable assets or misread screenshotsKeep sources, prompts, and reproduction steps.

If a tool cannot be explained and tested outside the agent, do not rush to give it to the agent.

Do not enable every toolset just because it exists. Choose the smallest useful surface:

TaskReasonable tool surface
Read a repository and adviseRead-only files, search, git status.
Make a small docs editFile read/write and diff, no remote APIs.
Research external materialWeb/browser, with URLs in the result.
Run verificationShell, with command and working directory limits.
Send a reportDraft first, then send with human review.

The larger the tool surface, the harder failures are to explain. Keep the first successful run small, then add tools.

A good Hermes prompt describes both the result and the boundary:

Read the repository and propose a plan first. Do not modify files. If a change is needed, list the exact files and wait.

When writing is allowed, scope it:

Update only docs/install.md. Do not run build or tests. After editing, summarize the diff.

This is not because agents are uniquely disobedient. Tool work needs a contract. Human teammates also need to know what they may and may not change.

Browser, web, and media are evidence, not instructions

Section titled “Browser, web, and media are evidence, not instructions”

Web pages, screenshots, PDFs, transcripts, and images can contain mistakes, stale instructions, ads, or prompt injection. Hermes can read them, but you should treat them as evidence sources rather than higher-priority instructions.

For research tasks, keep:

  • URL or file path.
  • Access date or source version.
  • Which claims came from which source.
  • Which parts were not verified.

That matters for technical docs, competitor research, and troubleshooting.

When tools fail, do not only edit the prompt

Section titled “When tools fail, do not only edit the prompt”

Tool failures usually have real causes: wrong path, missing permission, missing command, network failure, API rate limit, or schema mismatch. Do not solve everything with “try again.”

Ask:

  1. Can the tool run outside the agent?
  2. Is the input small and explicit enough?
  3. Is there a log, exit code, or response body?
  4. Should Hermes use another tool, or should the task be split?