Gantry · stats

Corpus

milestone era - on/after 2026-06-18

Codebases

18

distinct projects measured

Runs archived

231

198 milestones · 23 sprints · 9 maps · 1 goal

Agent sessions

11,283

8,663 execute · 384 plan · 1,692 review · 99 fix · 287 replan

Output tokens

57.5M

across every measured session

File edits

25,335

across every measured session

Execute-agent peak context

1,627 build sessions - exact per-session peaks

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 1,627 autonomous execute sessions, the peak context reached stays under the 150k design target in 80% of them - and 92.4% stay under 200k. Only 0.5% of sessions (8 of 1,627) peaked over 300k.

Under 150k target

80%

1,296 of 1,627 sessions

Under 200k

92%

only 124 sessions above

Median peak

104k

across all execute sessions

Over 300k

0.5%

8 sessions

claude execute 768 sessions

0

25

50

75

100

125

150

175

200

225

250

275

300+

150k 200k

74%< 150k

113kmedian

11%> 200k

1%> 300k

530kmax

codex execute 859 sessions

0

25

50

75

100

125

150

175

200

225

250

275

300+

150k 200k

85%< 150k

99kmedian

5%> 200k

0%> 300k

245kmax

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
execute	112,640	768	99,208	859
plan	83,035	187	45,556	197
review	62,717	775	68,999	912
fix	63,723	23	54,418	64
replan	53,736	167	40,923	120
milestones	41,599	71	27,800	87

Median over all roles: 73,399. claude peaks higher than codex in every role except review.

Decomposition

221 structured builds

Gantry sequences work by modelling it as milestones and sprints, and dispatching agents to implement sprint by sprint, and dispatching agents to do planning on multiple levels. Here is how the milestone-sprint metaphor ends up slicing the work.

Sprints / milestone

4.3

median - p25-p75 3.5-5.0

Milestones / job

3

median of the 108 jobs that split into milestones

Sprints / job

6

median across all 221 builds - max 54

How a plan breaks down

Metric	min	median	mean	p90	max
Milestones / multi-ms job n=108	2	3	2.8	4	8
Sprints / job n=221	1	6	8.1	16	54
Sprints / milestone n=108	1.3	4.3	4.4	5.7	9.0

Milestones per job

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

flat 113

2 ms 50

3 ms 37

4 ms 14

5 ms 5

6+ ms 2

Sprints per job

221 structured builds, bucketed by total sprint count.

1–2 sp 32

3–5 sp 77

6–10 sp 56

11–15 sp 28

16–25 sp 20

26–40 sp 7

41+ sp 1

Environment briefing

30 of 231 runs briefed

Gantry recently added a new environment briefing agent to the start of each milestone, in the hope that all subsequent agents will be better oriented, wasting less tokens searching project files. Only 207 briefed execute sessions so far across 30 of 231 jobs.

Execute peak - briefed vs not

Population	no brief	n	briefed	n	median delta
pooled	103,922	1,420	107,116	207	+3.1%
claude	109,733	734	164,759	34	+50.1%
codex	97,033	686	101,694	173	+4.8%