AI agent infrastructure moves from orchestration to control
An expanded first-half sample shows that infrastructure already dominates the agent ecosystem. The next shift is happening inside it: from general orchestration toward memory, observability, and reliability.
Not a new infrastructure wave, but specialization inside an established layer
The infrastructure share barely moved: 84.4% in Q1 and 85.5% in Q2. Yet memory and context gained 1.7 pp, observability and evaluation gained 1.0 pp, and recovery and reliability gained 0.4 pp, while general orchestration declined by 2.2 pp.
Sample scale and maturity
Quarterly cohorts after repository deduplication.
Q1 → Q2
95% CI: 83.8–87.1%
top 100 per quarter
exact project URL match
What is changing inside the stack
Each project belongs to one primary layer to avoid double counting.
Orchestration remains the largest segment, but its relative share is shrinking. Growth in narrower control functions points from assembling agent workflows toward operating them in long-running and critical processes.
Three market implications
Memory becomes infrastructure
Context, retrieval, and long-term memory are moving from application features into a reusable layer of agent systems.
Quality requires observability
Growth in evals, tracing, and monitoring reflects a new bottleneck: agent value depends on reproducible production behavior, not demos.
Reliability is small but accelerating
Recovery represents only 1.2%, but it is growing from a small base. This is an early market, not an established one.
Supporting evidence
These sources support interpretation but do not replace the primary GitHub statistics.
Signals tested by this research
A direct link between AILANTA's early observations and the study's measured findings.
GitHub: new public repositories from January–June 2026, combining stored broad cohorts with targeted searches for agent, agentic, MCP, multi-agent, computer-use, and tool-calling. Repositories are deduplicated by ID. Classification uses names, descriptions, and topics. Maturity is checked only for the top 100 agent projects in each quarter. Product Hunt is excluded.
Limitations: stars introduce survivorship bias; OpenAlex searches overlap and may lag; PyPI verification requires an exact GitHub project link; the official Show HN API only provides a current snapshot. The change in overall infrastructure share sits within overlapping confidence intervals and does not by itself establish growth.