CLAUDE CODE AGENTS: ADVANCED ORCHESTRATION - Chapter 7, Exercise 3
Why "More Shared Context Can Only Help" Is Flawed
====================================================================================

QUESTION:
A team, having learned about shared state, decides to add every single
detail from every agent's output into it, reasoning "more shared context
can only help." Using this chapter's own warning box, explain what's wrong
with this approach.

SOLUTION / EXPLANATION:

This chapter's warning box is explicit that indiscriminately dumping
everything from every step into shared state recreates the exact same
context-bloat problem Fundamentals' Chapter 1 identified for a human's own
conversation - it doesn't solve the noise problem, it just relocates it
into the agent chain's own shared state instead.

The team's reasoning treats all information as equally valuable simply
because it exists, but this chapter distinguishes genuinely worth
preserving information (the original task's hard constraints, key
decisions actually made) from disposable, step-specific working detail
that later agents in the chain don't actually need. If every intermediate
observation, exploratory dead-end, and minor detail from every single
agent gets added to shared state without distinction, the shared state
document itself grows into exactly the kind of unfiltered, noisy pile this
chapter warns against - and every later agent reading it now has to wade
through that bloat to find whatever's actually relevant to its own task,
which is a real cost, not a neutral non-issue.

So "more shared context can only help" is specifically wrong because
shared context isn't free to consume - a later agent still has to process
whatever's in shared state, and a bloated, undifferentiated dump makes that
harder, not easier, even though nothing was technically lost. The chapter's
own guidance is to be deliberate about what's added: genuinely important,
chain-wide information (like hard constraints) belongs there; a given
step's own disposable working detail generally doesn't, precisely because
indiscriminate inclusion defeats shared state's own purpose of keeping what
matters clearly visible.

--------------------------------------------------------------------------
WHY THIS WORKS AS AN ANSWER:
It explains specifically why "more shared context" isn't automatically
beneficial (later agents still have to process it, and bloat makes that
harder), and ties this directly to the chapter's own explicit distinction
between genuinely chain-wide information and disposable step-specific
detail.