The agent wave is shifting from brands to practice
An expanded sample of new GitHub projects confirms the scale of agent development while revealing a more nuanced shift: the broad layer is stable, branded waves are fading, and portable skills and instruction quality are gaining share.
Not continuous growth, but normalization after a March peak
ai-agent rose from 10.4% in January to 15.4% in March, then normalized to 11.2% in June—still above January. The ecosystem is not contracting; it is becoming less dependent on a single brand or protocol.
Six months of agent development
Author-provided topic share across all 695 repositories in each month.
Topics changing share
Three persistent shifts
Prompt engineering is rising again
Share rose from 0.7% to 2.4%. As agents become more complex, instructions and context management return as infrastructure problems.
Agent skills persist
agent-skill moved from 4.2% to 4.9% despite volatility, supporting a shift toward portable capabilities.
Branded themes lose concentration
Claude Code, Claude, MCP, and OpenClaw all lost share. Their use cases persist but diffuse across a broader tool stack.
Topic coverage
Cohorts are equal in size; metadata coverage is reported separately.