Stats

The creator of Gantry collects statistics on his own Gantry jobs and presents them here, to give an idea about how Gantry performs in real-life situations.

2026-06-18 - 2026-09-20

28 codebases 473 runs archived

877 milestones 4.5k planned tasks

20k agent sessions

930 plan 4.7k build 154 fix 437 of 473 runs

136.4M tokens out 66k edits

359k messages 385k commands 437 of 473 runs 435 of 473 runs 437 of 473 runs 435 of 473 runs

Execute-agent peak context

4,464 build sessions - exact per-session peaks 415 of 450 runs

The main goal of Gantry is to break work down into chunks small enough for agents to execute without running up large context windows. The key metrics are about context windows, and we focus on the peak number: the largest size the context window reached at any time during the session.

Across 4,464 autonomous execute sessions, the peak context reached stays under the 150k design target in 71% of them - and 88% stay under 200k. Only 0.6% of sessions (25 of 4,464) peaked over 300k.

Under 150k target

71%

3,175 of 4,464 sessions 415 of 450 runs

Under 200k

88%

only 534 sessions above 415 of 450 runs

Median peak

117k

across all execute sessions 415 of 450 runs

Over 300k

0.6%

25 sessions 415 of 450 runs

claude execute 987 sessions 112 of 450 runs

100

125

150

175

200

225

250

275

300+

150k 200k

68%< 150k

121kmedian

16%> 200k

2.5%> 300k

542kmax

codex execute 3,477 sessions 309 of 450 runs

100

125

150

175

200

225

250

275

300+

150k 200k

72%< 150k

117kmedian

11%> 200k

0%> 300k

246kmax

Each histogram is that harness's own execute sessions, bin width 25k, final bar 300k+; bar height is scaled to the busiest bin in that harness.

Supporting roles

the aside - non-execute agents

Execution is the most important type of agent run and the most context-constrained. The bottleneck we are targeting. Here are the other roles and their context peaks.

Per-role median by harness

Role	claude	n	codex	n
environment-build 101 of 450 runs	109,110	5	76,650	360
plan 360 of 450 runs	74,576	304	40,155	850
gate-build 101 of 450 runs	76,502	6	60,747	386
execute 415 of 450 runs	120,627	987	116,521	3,477
troubleshoot 124 of 450 runs	64,152	35	62,180	231
review 411 of 450 runs	64,618	1,028	75,910	4,212
replan 168 of 450 runs	55,289	203	40,577	214
resolve 34 of 450 runs	143,164	4	84,412	59
support 91 of 450 runs	36,895	2	27,950	134

Median over all roles: 79,631 416 of 450 runs. claude peaks higher than codex in every role except review.

Agent runtime

active working time per run and agent runtime by role

How long runs and agents take. The whole-run figure is active time: stopped time is excluded, and parked time is excluded where the journal records it. The harness columns are each harness's median; median, mean and p90 pool both.

Runtime per agent role

Role	claude	codex	median	mean	p90	n
environment-build 101 of 450 runs	5m 48s	2m 20s	2m 21s	2m 27s	3m 21s	364
plan 101 of 450 runs	3m 51s	1m 49s	1m 50s	1m 58s	2m 50s	401
gate-build 101 of 450 runs	5m 8s	2m 50s	2m 51s	3m 23s	6m 4s	392
execute 106 of 450 runs	16m 24s	6m 23s	6m 25s	8m 8s	14m 17s	1,505
troubleshoot 45 of 450 runs	5m 41s	1m 50s	1m 52s	3m 7s	5m 41s	117
review 101 of 450 runs	3m 35s	2m 6s	2m 6s	2m 40s	5m 7s	2,002
resolve 23 of 450 runs	10m 38s	4m 18s	5m 7s	6m 59s	11m 34s	29
support 91 of 450 runs	32s	32s	32s	36s	49s	153

Task runtime median

11m 25s

4,519 executed task rows - p90 29m 54s

Task runtime p25-p75

7m 31s - 18m 30s

execute + test + review per task

Active working median

2h 9m

whole-run active time

Active working mean

4h 24m

average across measured runs

Active working p90

8h 55m

90th percentile active time

Active working sample

473

runs with active duration

Stopped human waits are excluded for every active-duration observation. Parked retry waits are excluded for 7 newer runs; 466 older runs have stop-corrected active time only because their journals do not record parked waits.

Decomposition

450 structured builds

Gantry sequences work by modelling it as milestones and tasks, dispatching agents to implement task by task while planning on multiple levels. Here is how that structure ends up slicing the work.

Tasks / milestone

4.0

median - p25-p75 3.3-5.0

Milestones / job

median of the 315 jobs that split into milestones

Tasks / job

median across all 450 builds - max 57

How a plan breaks down

Metric	min	median	mean	p90	max
Milestones / multi-ms job n=315	1	3	2.8	4	8
Tasks / job n=450	1	8	9.8	20	57
Tasks / milestone n=315	0.5	4.0	4.4	6.0	12.0

Milestones per job

450 structured builds, bucketed by milestone count. A flat build is one whose tasks are a single list, with no milestone layer above them.

flat 135 135 of 450 runs

1 ms 44 44 of 450 runs

2 ms 103 103 of 450 runs

3 ms 90 90 of 450 runs

4 ms 50 50 of 450 runs

5 ms 19 19 of 450 runs

6+ ms 9 9 of 450 runs

Tasks per job

450 structured builds, bucketed by total task count.

1–2 tasks 43 43 of 450 runs

3–5 tasks 126 126 of 450 runs

6–10 tasks 131 131 of 450 runs

11–15 tasks 68 68 of 450 runs

16–25 tasks 61 61 of 450 runs

26–40 tasks 17 17 of 450 runs

41+ tasks 4 4 of 450 runs

Gantry builds Gantry

this repo's own history, through 2026-09-19

Gantry's own repository - the orchestrator and this website - is largely built by Gantry runs. Gantry commits under its own git identity, so its share of the work is measurable: every figure here is read from the repo's commit history and the records its runs left behind.

Runs on this repo

255

216 ran to completion

Agent sessions

6,908

530h 5m of agent runtime

Code lines by Gantry

75.8%

1,533,875 of 2,023,569 changed lines

Commits by Gantry

66%

5,284 of 8,005 commits

Who changed what

Changed lines and files across the whole history, by author identity. Gantry's run records and vendored third-party trees answer nothing about who did the work, so the narrower cuts drop them.

Cut	Gantry files	Gantry lines	other files	other lines	Gantry share
everything	34,004	6,331,502	33,732	4,546,989	58.2%
project files	3,272	2,110,209	3,653	945,903	69%
code only	1,964	1,533,875	1,674	489,694	75.8%

"Other" is every commit not made under Gantry's identity - including inline agent sessions that commit as the human, so it means not-Gantry, not hand-written. File counts overlap: a file both sides touched counts once for each. Of the 1,180 code files in the tree today, 719 (61%) were last written by Gantry.