Long-running agent costs are becoming a systems problem
An eight-month implementation cohort measures the controls used to contain context growth, repeated execution, retries, and model spend in long-running AI agents.
Early evidence; the market is not yet confirmed
27 subject-filtered implementations show that agent cost control is becoming a runtime discipline. Token accounting appears in 11.1% of the cohort, context compaction in 55.6%, prompt caching in 7.4%, and task-level cost benchmarks in 3.7%. Tooling is addressing context growth, but comparable cost-per-success measurement remains thin.
Market snapshot
Comparable measurements from independent market surfaces.
Implementations and packages
Deduplicated and subject-filtered primary cohort.
Median repository stars
Calculated across repositories with at least one star.
Token and cost accounting
3 items in the subject-filtered cohort.
Task-level cost benchmarks
1 items in the subject-filtered cohort.
Search-match dynamics
Bars show monthly repository search matches. The detailed cohort is filtered for explicit cost or context-control implementation language; it measures visible tooling supply, not aggregate enterprise AI expenditure.
What exists inside the category
One item may contain more than one feature.
Token and cost accounting
3 · 11.1%
Prompt and context caching
2 · 7.4%
Context compaction
15 · 55.6%
Selective memory retrieval
2 · 7.4%
Cost-aware model routing
0 · 0%
Retry and recovery controls
2 · 7.4%
Task-level cost benchmarks
1 · 3.7%
Representative projects
Cross-source validation
These publications are not part of the primary numeric cohort.