RESULTS & DEMOS
A coding benchmark result, followed by six real task stories. Follow the capability changes and the products they help create.
01 / BENCHMARK
Resolved tasks · Higher is better
AGENTEVOLVER
Formal evaluation on an external machine.
Result supplied by the AgentEvolver team.
AgentEvolver: team-reported. Reference scores: CodingFleet snapshot, updated Sep 11, 2026. Models, harnesses and budgets differ; this is a reported-score comparison, not a controlled comparison or an official ranking.
The team supplied 82.08% for SWE-bench Pro Public from a separate evaluation machine. The underlying model, run configuration and raw report are not included with this page. The local interrupted experiment is not the source of this score.
Reference values are transcribed from the linked compilation; their upstream links are retained in the downloadable data. The selection is not exhaustive. No result here measures the isolated contribution of self-evolution.
02 / EVOLUTION IN PRACTICE
4 websites · 1 game · 1 research workbench
80 seconds each · 1080p · English & Chinese editions
Discover a gap → evolve a capability → evaluate it → use it again. Each film connects that process with the actual product.
EN · 01:20Budget-aware asset retrieval, route design and paired play review evolve together.
666,655 bytes of verified assets within a 700,000-byte budget.
Self-play evidence; no independent player study.
EN · 01:20Semantic revision auditing grows into reusable dependency tracing.
4 of 9 objects flagged for review, with links to affected actions.
A demonstrated project case; no speedup claim.
EN · 01:20Source retrieval, a route-design agent and a review skill connect into one workflow.
The new review method detects 5 issues missed by a structural check.
Museum experiment counts are teaching examples, not platform benchmark scores.
EN · 01:20Ephemeris validation and investigation design improve alongside the 3D observatory.
Rejects mismatched dates; the updated connector is reused for Mars data.
The review skill was not adopted; the full evolution target remained incomplete.
EN · 01:20Luminance analysis and native font probes improve lighting and verify actual font rendering.
30.1% less median image-analysis time on the same screenshot.
Tool timing, not game FPS. The delivered game is a playable prototype.
EN · 01:20Data access, factor evaluation and strategy environments evolve into a traceable workbench.
77 evaluated factors · 40 formal strategy versions · final equity and drawdown.
The final strategy did not establish an advantage over its declared benchmark.