Fleet Active Time
68.4%
● Harness Layer · Operational Telemetry
Actor Activity & Session Yield
Observed actor time split into active execution, idle-with-work, unparked container hold, and leaf stall, with ceremony versus engineering calls and canonical session yield. A null yield is not a measured zero. Partitioned by actor and session across execution bands.
Each agent seat's observed window is partitioned into 5 discrete execution states across stacked horizontal bands:
Active Execution (Emerald)
Model is generating tokens, executing bash commands, or running tools. High active percentage indicates direct productive throughput.
Idle With Work (Amber)
Actionable friction: work is queued and assigned, but the agent process has not picked it up or is blocked on unpolled turn buffers.
Idle Without Work (Slate)
Benign rest: the assigned queue backlog is empty. The agent is parked waiting for human or supervisor dispatch.
Container Hold & Leaf Stall (Rose/Striped)
Process memory retained without execution, or leaf subprocess frozen waiting on parent synchronization.
Core performance equations and target operational thresholds:
Fleet Active Time
active_secs / total_observed_secs
Target: > 65% Healthy
Idle With Work (Friction)
idle_with_work_secs / total_observed_secs
Warning if > 15%
Ceremony Ratio
ceremony_calls / (ceremony_calls + tool_calls)
Warning if > 12% overhead
M1 Yield
tokens_burned / closed_nonreturned_items
Lower tok/item is better
Idle With Work (Actionable)
14.2%
Unparked Container Holds
450s
Fleet Ceremony Ratio
11.8%
Fleet M1 Yield
84.2K
Actor Activity Duration Bands
Aggregate observed duration per actor partitioned into discrete execution bands
Active Execution
Idle with Work
Idle without Work
Container Hold
Leaf Stall
Actor Idle & Yield Ranking
0 Actors
| Actor ID | Observed (s) | Active (s) | Idle w/ Work (s) ▾ | Idle w/o Work (s) | Container Hold (s) | Leaf Stall (s) | Ceremony Ratio | M1 Yield (tok/item) |
|---|
Technical Lineage, Coverage & SQL Proof (L3)
Underlying SQL Projection
SELECT * FROM v_seat_idle_ranking ORDER BY idle_with_work_secs DESC, seat_id;
Connected fact rows
Recorded window: —
Total rows: —