This is a genuinely good question and I want to be honest about what I think rather than either flattering or dismissing the work.

My honest assessment (~0.70 confidence):

The core algorithmic contribution of Hyperagents is not particularly complex in the traditional CS theory sense. Let me break down what's actually novel:

The DGM (the predecessor) is: maintain a population of agents, use an LLM to generate code mutations, evaluate them, keep the good ones in an archive, repeat. That's evolutionary search with LLM-guided mutation instead of random mutation. The archive with diversity incentives is basically MAP-Elites or Quality-Diversity, which has been around since ~2015.

What Hyperagents adds: make the meta-level (the mutation-generation procedure) also editable. Instead of a fixed function that generates mutations, the mutation-generator is part of the same codebase that gets mutated. That's... one design decision. It's an important one! But it's not algorithmically deep in the way that, say, FlashAttention's tiling strategy is, or the way optimal transport theory underpins Wasserstein GANs, or the way Transformers required understanding attention as a kernel method.

Compare to papers that require more raw algorithmic sophistication:

• FlashAttention (Tri Dao): requires understanding GPU memory hierarchies, SRAM/HBM tradeoffs, tiling strategies, IO-complexity analysis. The algorithm is genuinely novel and requires deep knowledge of both theoretical CS and hardware architecture simultaneously.
• Sparse attention mechanisms (various): require understanding approximation theory, random projection, locality-sensitive hashing.
• Formal verification work (like CompCert, seL4): requires understanding type theory, operational semantics, dependent types, proof assistants. Years of painstaking formal reasoning.
• Koppel's meta-metaprogramming (which we've discussed extensively): requires understanding tree automata, operational semantics, program transformation theory, the Curry-Howard correspondence. Genuinely deep PL theory.
• Anything involving cryptographic proofs: zero-knowledge proofs, homomorphic encryption — these require sustained mathematical reasoning at a level that Hyperagents simply doesn't.

What Hyperagents requires instead:

The intellectual contribution is more in the framing and experimental design than in algorithmic novelty. Specifically:

1. The insight that task-solving and self-improvement can be unified into a single editable program — this is a design pattern, not an algorithm. It's clever. It's not mathematically deep.
2. If the evaluation task were instead poetry writing, improving an agent's poetry-writing ability would not necessarily improve its ability to modify its own code — identifying this limitation of the DGM and seeing how to fix it. That's good taste in research direction. But the fix (make the meta-level editable too) is almost obvious once you state the problem.
3. The experimental work across four domains with proper ablations, transfer experiments, and analysis of emergent behaviors. This is where the effort is. Running these experiments, analyzing the code changes, documenting what the system spontaneously invented (performance tracking, persistent memory, compute-aware planning). This is thorough empirical science. It's not trivial — it takes engineering skill, patience, and good experimental design.
4. The connection to the Stanley/Clune open-endedness program — situating this within a 15-year intellectual tradition. Zhang understands the lineage and extends it meaningfully.

So: raw IQ breakdown (rough, ordinal, not precise):