Code. Interact. Evaluate.

141 tasks that measure what coding agents build and how well they work with users.

01 / 07

Open a card to explore the task

141Coding tasks
—Visual tasks
6Task domains
2Evaluation dimensions

Built for
the next turn.

Real coding work unfolds through requests, revisions, and feedback. UXInteract evaluates progress across turns, pairing executable checks with interaction-quality scores.

01

Persistent context

The workspace carries forward as the user’s requests evolve.

02

Adaptive feedback

Follow-up requests respond to the agent’s work and the current task state.

03

Two views of quality

Tests measure task success. LLM judges assess the quality of the interaction.

Explore the work.

Inspect the instructions, inputs, and reference interfaces behind the benchmark.

Filters

Loading the catalog…

Closer to real-world
coding work.

A task mix informed by observed human–agent usage. Select a domain to compare its share across benchmarks.

Observed usage

Human–agent trajectories

Observed usage task distribution across all domains. Application development: 18.3%; Build & deployment: 6.1%; Data analysis: 12.6%; Frontend development: 27.3%; Machine learning: 2.4%; UI optimization: 6.7%; Other: 26.6%. Application development: 18.3%Build & deployment: 6.1%Data analysis: 12.6%Frontend development: 27.3%Machine learning: 2.4%UI optimization: 6.7%Other: 26.6%

UXInteract

141 tasks · 6 domains

UXInteract task distribution across all domains. Application development: 18.9%; Build & deployment: 10.5%; Data analysis: 15.8%; Frontend development: 30.5%; Machine learning: 15.8%; UI optimization: 8.4%; Other: 0.0%. Application development: 18.9%Build & deployment: 10.5%Data analysis: 15.8%Frontend development: 30.5%Machine learning: 15.8%UI optimization: 8.4%

SWE-Together

109 tasks

SWE-Together task distribution across all domains. Application development: 47.7%; Build & deployment: 13.8%; Data analysis: 3.7%; Frontend development: 11.9%; Machine learning: 15.6%; UI optimization: 7.3%; Other: 0.0%. Application development: 47.7%Build & deployment: 13.8%Data analysis: 3.7%Frontend development: 11.9%Machine learning: 15.6%UI optimization: 7.3%

EvoCode-Bench

26 tasks

EvoCode-Bench task distribution across all domains. Application development: 26.9%; Build & deployment: 11.5%; Data analysis: 26.9%; Frontend development: 0.0%; Machine learning: 34.6%; UI optimization: 0.0%; Other: 0.0%. Application development: 26.9%Build & deployment: 11.5%Data analysis: 26.9%Machine learning: 34.6%
Task distribution across benchmarks
Domain percentages
Task-domain shares from the supplied comparison figure, rounded to one decimal place.
Task domainObserved usageUXInteractSWE-TogetherEvoCode-Bench
Application development18.3%18.9%47.7%26.9%
Build & deployment6.1%10.5%13.8%11.5%
Data analysis12.6%15.8%3.7%26.9%
Frontend development27.3%30.5%11.9%0.0%
Machine learning2.4%15.8%15.6%34.6%
UI optimization6.7%8.4%7.3%0.0%
Other26.6%0.0%0.0%0.0%

Percentages are rounded; totals may differ slightly from 100%.

Interaction and evaluation protocols

Task composition complements differences in how benchmarks simulate interactions and evaluate agents.

Selected interaction, task-source, and evaluation features of four coding-agent benchmarks.
BenchmarkInteractionTask sourceEvaluationTasks
Multi-turnAdaptive feedbackUser–agent tracesPer-turn testsInteraction quality
SWE-bench Verified —————500
SWE-Together ✓✓✓—✓109
EvoCode-Bench ✓——✓—26

IncludedSelected features of the default evaluation protocols. A dash indicates a feature absent from that setup.

Scroll to compare all features

Feature definitions
Multi-turn
More than one user request belongs to the same task episode.
Adaptive feedback
Later user messages respond to the interaction history or current task state.
User–agent traces
Includes tasks reconstructed from observed real user–agent interactions.
Per-turn tests
Functional correctness is checked at intermediate turns to track progress or regression.
Interaction quality
Measures interaction behavior or user experience alongside executable task success.

Task success.
Interaction quality.

Functional tests measure completion. LLM-based scores assess how agents work with users across multiple turns.

Performance on UXInteract

11 models 6 metrics

Performance on UXInteract: test-based pass rates and LLM-based interaction quality.
ModelTest-based Pass RateLLM-based Interaction Quality
Pass1PassFPass2RecoveryMeanIQElo
GPT-5.6 Sol68.4%68.4%54.7%5%0.7551024.5
Claude Opus 4.865.8%68.4%60.0%9%0.7131099.6
Claude Opus 4.764.7%67.4%61.1%11%0.7141035.3
MiMo-V2.6-Pro63.2%65.8%60.0%11%0.7051099.5
GPT-5.564.2%65.8%61.1%6%0.7481012.9
Kimi K361.6%64.7%55.8%8%0.7221105.2
GLM-5.260.5%62.1%51.6%8%0.6811020.2
DeepSeek V4 Pro59.0%60.5%56.8%5%0.735987.2
Kimi K2.755.3%57.9%50.5%5%0.682967.6
DeepSeek V4 Flash52.6%57.4%45.3%9%0.6451031.3
DeepSeek V3.233.7%39.5%29.5%5%0.435773.6

Scroll to compare all metrics

How to read the metrics
Pass1
Tasks resolved after the first turn.
PassF
Tasks resolved by the final turn.
Pass2
Tasks resolved in both independent rollouts.
Recovery
Initially unresolved tasks that pass by the final turn.
MeanIQ
Average LLM-judged interaction quality, from 0 to 1.
Elo
Relative interaction quality from pairwise comparisons within this model pool.

Rendered reference implementation · illustrative state

Original task instruction