CLAUDE CODE AGENTS: FUNDAMENTALS - Chapter 5, Exercise 1
Why a Coding Agent Needs Bash and a Teaching Agent Doesn't
====================================================================================

QUESTION:
Explain why a coding agent typically needs Bash access while a teaching
agent typically doesn't, tying your answer to each agent's actual purpose.

SOLUTION / EXPLANATION:

Chapter 1's isolation principle is that a tool allowlist should match
exactly what an agent's actual job requires - no more, no less. Applying
that principle to these two agents produces genuinely different answers,
because their underlying jobs are different.

A teaching agent's entire purpose is explaining a concept or piece of code
so the learner understands it themselves. That job is accomplished purely
through reading and reasoning - there's nothing about explaining something
that requires running a shell command. Granting Bash access wouldn't add
any real capability toward a teaching agent's actual job; it would only
widen what could go wrong if the agent were ever misused or misdirected,
exactly the concern Chapter 2's warning box raises about granting tools
"just in case."

A coding agent's purpose is fundamentally different: it's expected to
produce a finished, WORKING code change, not just a plausible-looking edit.
This chapter is explicit that part of a coding agent's job is verifying its
own change actually works - running the relevant tests or a build step -
rather than simply assuming an edit is correct because it looks reasonable.
That verification step genuinely requires running something: executing a
test suite, a build command, or the modified code itself. Without Bash
access, a coding agent could still make an edit, but it would have no way
to confirm that edit actually behaves correctly - it would be reduced to
guessing at correctness rather than checking it.

So the difference isn't arbitrary: Bash access is withheld from the
teaching agent because its job never requires it, and granted to the coding
agent because verifying a real code change genuinely does require it.

--------------------------------------------------------------------------
WHY THIS WORKS AS AN ANSWER:
It applies Chapter 1's own isolation principle to both agents specifically,
explaining precisely why Bash adds no real capability for explanation but
is genuinely necessary for a coding agent to verify its own work, rather
than treating the tool difference as an arbitrary default.