Novel Problems vs Recombined Ones: The Line That Matters

A useful predictor of whether an agent will handle something well: is this problem a new arrangement of solved pieces, or does it require an approach that doesn't exist yet in any form?

Almost all engineering is the first kind. That's why agents are as useful as they are — and identifying the minority that isn't tells you where to spend human attention.

What recombination covers

Most work is a novel arrangement of familiar components. A new API endpoint is a known shape applied to new fields. A caching layer is a known pattern applied to a specific access profile. A migration is a known procedure applied to your schema.

The arrangement may be genuinely new — nobody has built exactly your system — while every element is well established. Agents are strong here for a straightforward reason: the pieces are represented, the composition rules are represented, and the result is checkable against tests.

⚠️ Recombination is not "easy." Composing familiar elements correctly, in a large system, under real constraints, is most of what makes engineering hard. Calling it recombination isn't a slight; it's a statement about where the pieces come from.

What genuinely novel looks like

Rarer than people think, and identifiable by a specific property: there's no existing approach to adapt.

• A performance problem whose cause is an interaction nobody has documented, in a combination of systems nobody else runs.
• A domain constraint with no analogue — a regulatory rule, a physical process, a business model that doesn't map onto known patterns.
• A failure mode arising from your specific composition of components, where each behaves as documented and the combination doesn't.
• A problem where the standard approach is known and doesn't work here, for reasons specific to your situation.

The tell is that searching produces nothing that fits. Not "nothing exact" — nothing whose shape can be adapted.

Why agents fall off at the boundary

Not that they can't recombine cleverly — they can, at wider scope than most people. The failure is subtler.

Confident near-misses. Given a novel problem, an agent retrieves the nearest known pattern and applies it. The result is a solution to an adjacent problem, presented with the same confidence as a correct one. This is worse than an admission of difficulty, because it looks like progress.

No signal that the ground changed. Nothing announces "this is outside what I've seen." The output for a novel problem looks like the output for a familiar one.

→ The practical consequence: on genuinely novel problems you don't get an error, you get a plausible wrong answer. The burden of noticing sits entirely with the human.

🔍 Telling them apart

Before handing something over:

• Does a name exist for this problem? Named problems have known solutions. Unnamed ones may not.
• Would searching find something adaptable? If yes, recombination. If everything found is adjacent-but-not-applicable, be careful.
• Does the standard approach fail here, and do you know why? That's the signature of novelty — and knowing why it fails is usually most of the solution.
• Is the difficulty in the composition or in the concept? Composition difficulty is agent-friendly. Conceptual difficulty isn't.

✅ Using agents on novel problems anyway

They're still useful, in a different mode:

• Enumerate what's been tried elsewhere, so you know what you're ruling out. Recall is the strength.
• Explore the adjacent space — problems that are nearly this one, and what they do. Sometimes the adaptation is visible once the neighbours are laid out.
• Implement the novel approach once you've designed it. The design is the novel part; the implementation is recombination again.
• Check your reasoning, as a skeptical reader looking for holes.

What doesn't work is asking for the answer, because the answer doesn't exist to be retrieved.

💡 Why this matters for how you spend attention

If most work is recombination and agents handle recombination well, then human attention should concentrate on the minority that isn't — plus the judgment about which is which.

That classification is itself a skill, and it's the one most likely to go wrong quietly. Misclassifying a novel problem as routine means accepting a confident wrong answer. Misclassifying routine as novel means doing by hand what could have been delegated — a cost, but a recoverable one.

The asymmetry says: when unsure, treat it as novel and stay involved.

The takeaway

Nearly all engineering is recombination of solved pieces, which is precisely what agents do well and why they're useful. The minority requiring an approach that doesn't yet exist gets confident near-misses instead of errors. Classify before delegating — does a name exist, would searching find something adaptable, does the standard approach fail and do you know why — and when the classification is uncertain, assume novel and stay in the loop.