Structured Decomposition

Structured decomposition is an engineering agent pattern where a long-horizon generation problem is split into a pipeline of narrow agents connected by typed intermediate representations, with a deterministic checkpoint at every handoff. No single model owns the workflow end-to-end; each agent produces a bounded, validatable artifact.

The strongest current evidence comes from the University of Miami / HBC Engineering structural analysis paper (arXiv 2606.06525): a multi-agent pipeline converts structured natural-language briefs of irregular 3D frames into executable SAP2000 models at 90% mean accuracy (response agreement within 1% of reference models), while direct single-model generation with GPT-5.4 and Gemini-3.1 Pro scores 0% under the same test.

Why decomposition is the mechanism, not a preference

Complete engineering scripts are a long-horizon generation problem: small errors accumulate across thousands of lines that must preserve spatial meaning, shared topology, and software-specific syntax. The ablation results show each architectural element carries the load:

AblationEffect
Remove floor decomposition (intermediate representation)Accuracy drops to 20–50%
Merge node/girder/slab agents into oneAccuracy drops to 50–70%
Merge translation and compilation stagesAccuracy drops to 0%

The decisive design choice is not the number of agents. It is the presence of explicit intermediate representations and deterministic acceptance criteria at each boundary.

The pattern, generalized

  1. Intent capture — extract requirements into a typed contract (geometry, materials, boundary conditions, loads, scope)
  2. Domain representation — convert the hard reasoning problem into a stable intermediate form the agent can reliably manipulate (here: a matrix-of-stories → per-floor occupancy maps)
  3. Narrow generation agents — parallelizable agents with single responsibilities (nodes, girders, slabs, columns)
  4. Deterministic checkpoints — reject duplicate nodes, undefined endpoints, invalid corners, disconnected geometry before downstream processing
  5. Staged translation and assembly — produce executable software commands in steps, never in one pass
  6. Engineer release authority — generated models remain draft artifacts until a licensed engineer verifies them

A secondary platform insight: the pipeline assigns different models by task capability (a larger model for spatial reasoning, a lighter model for code translation), keeping cost near $0.19 and runtime near 175 seconds per case.

Relationship to FEA-in-the-loop

FEAInTheLoop and structured decomposition are complementary halves of the same governance architecture:

  • Structured decomposition validates during generation — checkpoints block invalid intermediate artifacts before they propagate
  • FEA-in-the-loop validates after generation — simulation produces typed failure evidence that drives repair

Both reject the same failure mode: asking one general-purpose model to own an entire engineering workflow and trusting plausible-looking output. The structural analysis paper’s 0% direct-generation baselines independently corroborate the FEA paper’s finding that first-shot generation is not enough for industrial use.

Boundary

The 90% benchmark covers ten author-designed orthogonal frame cases with standardized loads and materials. Not yet demonstrated: nonorthogonal or curved geometry, dynamic/seismic/wind response, shear walls and bracing, requirement diversity, regulatory compliance, or model certification. Data and code are available only on request, limiting reproducibility. Keep licensed engineer sign-off mandatory; pilot in shadow mode against trusted reference models.