Today's take
The strongest signal is that agent capabilities are becoming assets that can accumulate over time. Self-learning skills, persistent memory and diagram tools are gaining attention together, while creative projects focus on tangible deliverables.
App usage points in a similar direction: Hermes Agent leads by tokens, OpenClaw tops the growth ranking, and coding agents occupy several leading positions. The dates differ: GitHub rankings are for 2026-10-10; model and app usage is for 2026-10-08. Usage covers only traffic routed through OpenRouter with public attribution; tokens measure usage, not people.
Open source
Three themes stand out across the Top 50: visual deliverables, reusable experience and persistent memory. Today's prominent projects turn agent capabilities into diagrams, creative output and session knowledge management; stars and ranking movements indicate attention rather than adoption.
- Visual deliverables lead: #1 storytold/artcraft gained 4,247 stars and moved from #3 to #1; #2 cathrynlavery/diagram-design gained 1,556 and climbed from #6 to #2. A creative engine and diagram-design skills leading together suggest interest in usable visual output alongside text generation.
- #3 tt-a1i/archify: Added 1,416 stars, climbed from #22 to #3, and reached 81,021 total stars. Turning ideas, plans and codebases into interactive diagrams extends agent output into interfaces people can navigate and discuss.
- #6 tigerless-labs/autoharness: Added 1,196 stars and jumped from #40 to #6. Its approach extracts skills from real sessions, updates them and prunes unused ones, treating engineering experience as a maintained asset.
- #21 vectorize-io/hindsight: Added 515 stars and climbed from #33 to #21. Its focus on agent memory that learns complements adaptive skills: persistent workflows need both retained context and improving execution practices.
Models
Volume leadership and growth leadership diverged: the largest model declined while several other models expanded quickly. All three leading entries carry a batch label, but the ranking alone does not establish their underlying workloads or app sources.
- DeepSeek V4.1 Flash (batch): Remained #1 at 5.38T tokens despite a 27% daily decline. #2 GLM 5.3 Flash (batch) processed 1.78T, up 6%; #3 GPT-6 Luna (batch) processed 1.14T, up 20%.
- Claude Haiku 5.5 (batch): Reached 320B tokens at #11, up 835%, making it the strongest growth signal outside the top ten. Subsequent absolute volume will help distinguish a sustained shift from a short spike.
- Muse Spark 1.3 Contributor: Ranked #10 with 351B tokens, up 57%; Muse Spark 1.3 processed 75.1B, up 124%. These are separate entries showing concurrent growth worth tracking.
Apps
Personal agents and coding tools occupy the leading usage positions, and the growth ranking reinforces that direction. Growth is measured against the average of the preceding three equal-length periods; percentage increases were not supplied.
- Hermes Agent: Led usage with 1.43T tokens. Its persistent memory and reusable-skills positioning aligns thematically with the open-source momentum around skills and memory.
- Coding agents: Claude Code processed 1.28T, Codex 597B and Kilo Code 558B tokens, ranking #2, #3 and #4. Codex and Claude Code also ranked #2 and #3 on the growth list.
- OpenClaw: Processed 496B tokens, ranking #5 by usage and #1 by growth. Personal-agent momentum makes persistent task execution worth following alongside coding workflows.
What to watch
Watch whether attention to skills and memory projects persists alongside growing usage of persistent agents. The two datasets support a shared direction, but do not establish that these repositories drove app growth or that particular apps caused the model increases.
