A team from the University of Science and Technology of China and Hong Kong Baptist University published EvolveNet on August 5, 2026—a framework that improves LLM agents by evolving the executable harness surrounding a fixed model across five benchmark domains, without retraining weights.
The harness is the executable program that constructs context, invokes tools, verifies results, and recovers from failure. EvolveNet argues this layer—not the base model—is the primary optimization target. Testing across text-to-SQL, data-science coding, competitive programming, software engineering, and agentic workflows showed improvement in all five, with largest gains in heterogeneous workloads.
Prior harness-evolution systems—Darwin Gödel Machine, Meta-Harness, GEPA—assume all execution experience flows to a single optimizer that evolves one harness sequentially. That works in labs. It breaks when experience comes from distinct users, organizations, and environments that cannot pool workload data. Most enterprise deployments fit this constraint. EvolveNet distributes the shared harness to data-local agent deployments, lets each evolve it against its own workload, then collects only the resulting program adaptations for aggregation. Raw workloads stay local.
| System | Optimizer Model | Experience Aggregation | Data Sharing Required | Suitable for Multi-Tenant/Siloed Deployments |
|---|---|---|---|---|
| Darwin Gödel Machine | Single optimizer | All experience → one sequential optimizer | Yes — workloads must be pooled | No |
| Meta-Harness | Single optimizer | All experience → one sequential optimizer | Yes — workloads must be pooled | No |
| GEPA | Single optimizer | All experience → one sequential optimizer | Yes — workloads must be pooled | No |
| EvolveNet | Distributed (data-local) | Local evolution → program diffs collected → aggregated centrally | No — only program adaptations shared | Yes |
Merging independently modified programs is not parameter averaging—adaptations conflict at the scope level. EvolveNet introduces scope-typed, evidence-guided program aggregation to resolve conflicts. Ablation results show gains come from composing adaptations across agents, not selecting among them. For multi-tenant platforms, federation of program diffs—not selection of the best variant—drives collective improvement.
A parallel paper (arXiv:2605.30621) reveals where EvolveNet pays off: harness-updating capability is flat across model tiers. Qwen3.5-9B updates yield gains comparable to Claude Opus 4.6. But harness-benefit is non-monotonic. Mid-tier models benefit most; weak-tier models fail to activate or follow updated artifacts; strong-tier models benefit less than mid-tier. For stack decisions: the evolver does not need to be your strongest model, but the agent solving the task does.
| Model Tier | Example | Harness-Updating Capability | Harness-Benefit Level | Key Failure Mode |
|---|---|---|---|---|
| Weak | — | Flat (comparable to others) | Low | Fails to activate or follow updated harness artifacts |
| Mid | Qwen3.5-9B | Flat (comparable to strong models) | High — benefits most | None identified |
| Strong | Claude Opus 4.6 | Flat (comparable to mid models) | Moderate — less than mid-tier | Diminishing returns; less room to gain |
Another related paper (arXiv:2605.27276) frames the constraint sharply. Harness-only systems improve "software-engineering hygiene—parsing, retries, dispatch"—and rarely deliver domain-specific reasoning the base model cannot produce. Production agents lose significant performance to brittle error recovery and poorly structured tool invocations long before hitting the ceiling of what the base model can reason. Fixing the harness first is correct.
Harness evolution ships as program diffs, not gradient steps. No GPU cluster, no retraining, no serving swap. For teams running agents on third-party API models—where weight updates are impossible—EvolveNet-style harness evolution is the entire optimization surface.
Instrument deployments to capture program adaptations per workload domain. Build the aggregation layer that composes them. Treat that pipeline as the primary lever for production improvement before scheduling a fine-tune.