T3 C ode · P erformance over time
Queued run start p90
All metrics
Turns completed in the window Agents served in the window Turn end to end p50 Turn end to end p95 Dispatch p50 Dispatch p95 First streamed item p50 Push latency p95 Thread switch: bounded open p50 Thread switch: stream synchronised p95 Shell stream synchronised p95 MCP thread read poll p95 MCP project list poll p95 HTTP bounded read poll p95 Session MCP call p95 Thread events per client per second Shell events per client per second Failed calls Calls still running after the grace period Main thread busy Main thread user CPU Main thread system CPU Process CPU, all threads Worker threads CPU Child processes CPU Child processes alive Event loop utilisation, mean Event loop utilisation, max Event loop delay p50 (preload histogram) Event loop delay p99 (preload histogram) Event loop delay max (preload histogram) Event loop delay max (server gauge) GC pause per second Seconds with a stall of 40 ms or more Seconds with a stall of 100 ms or more Seconds with a stall of 500 ms or more Main thread 1 s busy p95 Main thread run-queue wait Main thread minor faults Main thread voluntary context switches Main thread involuntary context switches Main thread read syscalls Main thread write syscalls Process read syscalls Process CPU per served agent Peak RSS Peak V8 heap used Heap after the last GC Memory of every T3 process (PSS) Memory of every T3 process (USS) RSS slope over 2 h Heap slope over 2 h WAL size, max WAL size, median Reader transaction, mean Reader transaction p95 Reader lock wait, mean Reader lock wait p95 Reader statement p95 Writer transaction, mean Writer transaction p95 Writer lock wait p95 Reader transactions per second Writer transactions per second Reader slots busy, mean share Free reader while a request waited SQL statement time per second Thread shell read p95 Thread snapshot window p95 MCP request authentication p95 Provider turn start p50 Run update handling p95 Subagent finalisation per child event p50 Effect queue wait p50 Effect queue wait p90 Items per provider commit Spans in the window Trace lines dropped Git commands per turn Git time per turn Process creations per turn Main-thread gap after descendant polls, growth Start to listening, first start Start to listening, warm start Start to port accept, first start Start to port accept, warm start Start to database ready, first start Start to database ready, warm start Migrations on the era fixture, chain total Per-migration seconds Peak extra disk during migrations SIGTERM to exit HTTP probe max during compaction Prune rate Full snapshot GET of the largest thread Full snapshot size of the largest thread Idle CPU, one idle agent Idle wakeups, one idle agent Memory after 30 min idle First turn after a cold start Queued run start p50 Queued run start p90 Queued run start max Run update consumer backlog, max Run update consumer drain time Delivery latency by run state p50 Live event buffer full warnings Warning log lines Error log lines Stream resynchronisations Valid attempts Attempts Host busy Host steal Host I/O wait Contaminated attempts Effect queue wait p95 (span attribute) Effect worker run p95 Projection decode p95 Effect admission changes
Queued run start p90 (lower is better)
–
No records yet.
Agents 1 50 100 150 200 250 Range 7d 14d 30d
When upstream merges a fix, the lines meet.
Upstream nightly Performance fork Noise band Range of attempts▲▼ green = better, ▲▼ red = worse, beyond the noise band
As agents grow. 90th percentile time for a queued run on an unrelated thread to start, overall, during a cancel burst and in the 2 min after (fork releases and the weekly lane).
Lower is better. No noise band measured yet, so no verdict.
Unit s
Better lower
Noise band none measured yet: no verdict
Recorded from 20 Oct 2026; before that the series is a gap, never a zero
Platforms Linux x64
Lane Fork releases only
Upstream builds Every upstream build
Source lagbench.json: queue start p90
Queued run start p90 (lower is better) at 1 to 250 agents, Linux x64
against
Records, 50 agents
Ladder as a table
Download this series No series file yet.