You have some ambitious plan. You know Claude Code is not going to one-shot it. After a while, you get a feel for how many sessions would be needed. Gantry is a tool for extending the long-running capacity of agent harnesses, such that a single Gantry session will orchestrate between 5 and 40 agent sessions worth of work.
The smallest useful Gantry job is a single milestone with a handful of sprints, say 3-5 of them. If the job is smaller than that, Gantry tends to split it into too many steps, so you end up wasting agent restarts on work that one agent could have handled. On the other end of the spectrum, Gantry seldom goes beyond 8 milestones, so 8 milestones with a handful of sprints each gives a kind of upper bound for the current version.
The way I think about it, Gantry can comfortably handle at least 40 'execution' agents, i.e. excluding agents doing planning, reviewing, debugging etc. Beyond that it still works but it starts to feel a bit more experimental.
If Gantry sets up a 5 milestone run with 30 sprints, that often feels like a sizable but manageable chunk of work. When something becomes an 8-milestone run, I often feel that I have gone beyond what I am meaningfully able to plan out and simultaneously hold in my head.
In other words, implementing plans larger than Gantry can readily handle is not really useful for me, because I am then bottlenecked by my own cognitive capacity. Maybe people who are smarter than me and able to keep more things in their mind at one time will be building more advanced systems which focus on orchestrating hundreds of agent sessions rather than dozens.
Here is a snapshot of some actual Gantry stats:
Keep in mind that the stats are slightly misleading, for the reason I just hinted at. I've been pushing the limits to see where Gantry stops working properly. I have run many Gantry jobs on small tasks which could have been handled by a single Claude Code or Codex session. I have also run some very big jobs, which Gantry has split into 50+ sprints but despite this, too much has been forced into many of the sprints.
Gantry has ways of dealing with this. For example, both executing and reviewing agents can flag a sprint as too big, and a reviewer or debugger agent can then do replanning. A sprint can be broken up into smaller ones and entire new milestones can be added. This kind of thing is normal in Gantry. But when I give it a massive, unreasonably ambitious plan, it ends up doing a lot more replanning than usual.
At one point I saw a milestone being replanned from 10 to 15 to 20 sprints. I took this as a sign that Gantry had been forced to bite off more than it could chew. But, again, this kind of thing only happens when I give Gantry a job so big that it feels a bit silly. I personally see no good reason to have a single instance of an agent orchestrator running autonomously for days on end. If I was not out to stress test Gantry, I would break any such plan into a number of ~24 hour runs, because even after 24 hours, the work of reviewing the work of the agents starts to get overwhelming (given my level of cognitive capacity, your brain size may vary).
At one point I spent about two weeks planning a huge Gantry run, rewriting a very complex application with a lot of novel features from scratch (my procedural music generator project, as mentioned in this post about what projects Gantry has worked on). Gantry ran for several days, got through it and finished up a successful rewrite. It was a triumph. But I would not have attempted that in a single autonomous run unless I was testing the limits of the orchestrator.
In my own work, I'm trying to achieve a good planning-execution cadence that fits my way of thinking. What feels best to me is often to spend 1-4 hours on focused analysis and planning work, then having agents execute over 5, 10 or 20 hours. Spending multiple weeks planning something without producing any code – and then waiting a week for the execution to finish – feels lopsided and not very useful. Hopefully I will laugh at that last statement once I transcend to higher levels of 'dark software factory' enlightenment.