Persistent context
The workspace carries forward as the user’s requests evolve.
141 tasks that measure what coding agents build and how well they work with users.
Open a card to explore the task
Real coding work unfolds through requests, revisions, and feedback. UXInteract evaluates progress across turns, pairing executable checks with interaction-quality scores.
The workspace carries forward as the user’s requests evolve.
Follow-up requests respond to the agent’s work and the current task state.
Tests measure task success. LLM judges assess the quality of the interaction.
Inspect the instructions, inputs, and reference interfaces behind the benchmark.
Loading the catalog…
Try a different keyword or clear the filters.
Reference previews come from solution outputs. Other tasks use a bundled input or code excerpt.
A task mix informed by observed human–agent usage. Select a domain to compare its share across benchmarks.
Human–agent trajectories
141 tasks · 6 domains
109 tasks
26 tasks
| Task domain | Observed usage | UXInteract | SWE-Together | EvoCode-Bench |
|---|---|---|---|---|
| Application development | 18.3% | 18.9% | 47.7% | 26.9% |
| Build & deployment | 6.1% | 10.5% | 13.8% | 11.5% |
| Data analysis | 12.6% | 15.8% | 3.7% | 26.9% |
| Frontend development | 27.3% | 30.5% | 11.9% | 0.0% |
| Machine learning | 2.4% | 15.8% | 15.6% | 34.6% |
| UI optimization | 6.7% | 8.4% | 7.3% | 0.0% |
| Other | 26.6% | 0.0% | 0.0% | 0.0% |
Percentages are rounded; totals may differ slightly from 100%.
Task composition complements differences in how benchmarks simulate interactions and evaluate agents.
| Benchmark | Interaction | Task source | Evaluation | Tasks | ||
|---|---|---|---|---|---|---|
| Multi-turn | Adaptive feedback | User–agent traces | Per-turn tests | Interaction quality | ||
| SWE-bench Verified | — | — | — | — | — | 500 |
| SWE-Together | ✓ | ✓ | ✓ | — | ✓ | 109 |
| EvoCode-Bench | ✓ | — | — | ✓ | — | 26 |
| UXInteract | ✓ | ✓ | ✓ | ✓ | ✓ | 139 |
IncludedSelected features of the default evaluation protocols. A dash indicates a feature absent from that setup.
Scroll to compare all features
Functional tests measure completion. LLM-based scores assess how agents work with users across multiple turns.
11 models 6 metrics
The task explorer includes 141 tasks. Historical model scores retain their original evaluation scope.
| Model | Test-based Pass Rate | LLM-based Interaction Quality | ||||
|---|---|---|---|---|---|---|
| Pass1 | PassF | Pass2 | Recovery | MeanIQ | Elo | |
| GPT-5.6 Sol | 68.4% | 68.4% | 54.7% | 5% | 0.755 | 1024.5 |
| Claude Opus 4.8 | 65.8% | 68.4% | 60.0% | 9% | 0.713 | 1099.6 |
| Claude Opus 4.7 | 64.7% | 67.4% | 61.1% | 11% | 0.714 | 1035.3 |
| MiMo-V2.6-Pro | 63.2% | 65.8% | 60.0% | 11% | 0.705 | 1099.5 |
| GPT-5.5 | 64.2% | 65.8% | 61.1% | 6% | 0.748 | 1012.9 |
| Kimi K3 | 61.6% | 64.7% | 55.8% | 8% | 0.722 | 1105.2 |
| GLM-5.2 | 60.5% | 62.1% | 51.6% | 8% | 0.681 | 1020.2 |
| DeepSeek V4 Pro | 59.0% | 60.5% | 56.8% | 5% | 0.735 | 987.2 |
| Kimi K2.7 | 55.3% | 57.9% | 50.5% | 5% | 0.682 | 967.6 |
| DeepSeek V4 Flash | 52.6% | 57.4% | 45.3% | 9% | 0.645 | 1031.3 |
| DeepSeek V3.2 | 33.7% | 39.5% | 29.5% | 5% | 0.435 | 773.6 |
Scroll to compare all metrics