Ask one agent for ten ideas and you get one idea ten times

You're at the start of something and you want options. So you ask your agent: "give me a few ways to approach this." It obliges — a clean numbered list, three or four approaches, each with its little pro-and-con. You skim, pick the one that reads best, and move on feeling like you weighed the space.

You didn't. You weighed one corner of it, described four ways.

01More options from one agent isn't more diversity

Here's the thing about asking a single agent to brainstorm: every option comes out of the same head. Same training, same instincts, same sense of what the "obvious" move is. When you say "give me alternatives," it spreads out a little — but around one center of gravity. Options two through four are usually option one wearing different clothes: the same core approach, reframed, re-scoped, lightly reskinned. It looks like a menu. It's one dish, plated four ways.

And here's the part that bites: whatever that mind doesn't reach for, none of the four will contain it. The blind spot isn't in one option — it's in all of them at once, because they share an author. You can't widen your view of the problem by asking the same pair of eyes to look harder.

02A second mind of a different make

This is the same move as bringing in a second set of eyes — only pulled all the way upstream, to the moment before anything is decided. Give the same problem to a second agent of a different make — Claude and Codex, say — and you get something a single mind can't produce on command: a second center of gravity. Where Claude reaches for the framework-shaped solution, Codex reaches for the script-shaped one. Where one assumes you'll need to scale, the other assumes you won't. Neither is right yet — but now you can see the assumption, precisely because the other one didn't make it.

The disagreement is the signal. When both land on the same approach, that's real evidence it's the obvious path — for better or worse. When they split, they've handed you a map of the actual decision: the fork you'd otherwise have walked straight past, never noticing there was a fork.

One rule makes or breaks this: ask them cold, and separately. The moment Codex sees Claude's answer first, it stops being a second mind and becomes a commenter on the first — it anchors, it agrees, it polishes. You wanted two independent draws from two different wells; the instant one reads the other, you're back to one well with an echo.

03The prompts: cast both the same, then diverge

None of this needs special tooling — two chat windows and the discipline to keep them apart. Cast each agent the same way first, so neither tries to hand you one polished answer:

You're one of two independent minds I'm asking to approach this separately. Don't try
to give me the single best answer — give me your genuine take and lean into what's
distinctive about it. I'll be comparing it against another agent's, so a different
angle is useful here, not a problem.

Then the task prompt, the same words to both, each cold:

Here's the problem: [goal in a sentence, the hard constraints, and what a good
solution would have]. Give me 2–3 genuinely different approaches — not variations
on one idea. For each: the core in two lines, its main trade-off, and the single
assumption it rests on. Don't rank them yet.

Collect both answers before either agent sees the other's — that order is the whole point. Then, and only then, the second prompt: paste one agent's approach into the other.

Here's a different approach to the same problem: [paste the other agent's answer].
Against your own: where is each stronger, where would each break first, and which
assumption do the two actually disagree on? Be specific about the disagreement.

What comes back from that second round is the real prize — not two lists, but a sharpened picture of the one or two decisions the whole thing actually turns on.

04Don't average them — harvest the fork

The failure at the end is as quiet as the one at the start. You get two good, different proposals and the temptation is to split the difference — take a bit of each, blend toward a compromise that offends no one. That usually gives you the worst of both: a design with two half-committed spines and no clean reasoning behind any single choice.

The value was never the average. It was the divergence — the specific places the two minds pulled apart, because each of those is a real decision you now get to make on purpose instead of by default. Pick a spine. Borrow the one genuinely better idea from the other. But decide — don't blend.

05Where the second mind pays for itself

It's overkill for "what should I name this function." It earns its keep at the decisions that are expensive to reverse — the architecture, the data model, the framing of the whole thing — where the locally-obvious path is exactly the one a single mind hands you first, and exactly the one you'll be living inside for months. Cheap to get a second draw now; costly to discover, a hundred commits later, that there was a fork you never saw.

The trap underneath all of this is a comfortable one: a fluent agent makes one corner of the solution space feel like the whole map. It isn't lying — it genuinely can't show you the part it can't see. A second mind of a different make isn't more hands on the same idea. It's the only cheap way to find out your best idea had a neighbor.