Depends on whether you can keep things separated logically. I have 3 git worktrees open, each working on a different area.
Generally its feature a, feature b and a refactoring branch of some kind.
My workflow is:
1. Add ticket in gitlab describing bug or feature in as much detail as possible along with acceptance criteria like expected unit tests or browser based tests.
2. In a work tree create a branch based on the id of that ticket in gitlab.
3. Start Claude, tell it to use a skill to pull that ticket, research and make a plan.
4. Review the plan, ask questions, refine.
5. Approve plan and let cluade cook.
6. Have Claude run a set of linters/tests/code quality checks and ground until done.
7. Start a new Claude instance, ask it to review changes made. Provide feedback to first Claude instance for changes.
8. Commit and push, creating a draft mr/pr in gitlab.
9. Review the actual code changes myself using gitlab. Comment on things not right.
10. Get Claude to use another skill to pull comments and work to resolve them. Also feed back any CI failures.
11. Manually close comments and push again. Repeat until done and ready for co-worker review.
I can only keep 3 threads like this going at once. Sometimes it’s only 1 or 2, depending on complexity. Smaller is better. Try to stay atomic and avoid feature creep in each mr.
Pretty much everyone saying worktrees but I lean heavily on hot-reloading of both backend and frontend (to actually see what I'm doing) so it's too annoying to deal with dependencies and ports when things are isolated. Instead I keep everything on the main branch and just make sure to keep tasks pretty separate in scope, e.g. add this API route vs. fix this layout issue, so they don't step on each other. When I consider something done I have the agent commit what it worked on and start a new one.
I do use worktrees occasionally (especially during times where I'll have a very sticky problem that I make the LLM run in a loop on until it satisfies acceptance criteria, and want to isolate the potential fallout of Claudes Gone Wild), and I run Claude and Codex side by side, but I rarely have them work on truly-different tasks simultaneously.
The main reason is because if there's a significant bug or large optimization going on, that shit needs to be done, tested and merged before building more stuff on top, otherwise you run the risk of wasted time, tokens and effort having a bunch of parallel work running that may not end up compatible at the end.
Lately I've had a lot more success having Claude generate a plan, send the plan to Codex for co-validation/amendments, have Claude implement the plan, then have Codex PR review the commit (and likely make some edits of its own), then I test out the code/changes.
Meanwhile, my actual management of what I'm asking them to do is just a text file in Notepad where I'll write like BUG: xyz thing does abc or IDEA: let's change this to that as I'm testing in-app, with the actual code opened in Notepad++ tabs (lol feel free to roast me, I'm in front of 2 screens, one Windows (primary), one Mac (to the right), sharing keyboard and mouse -- LLMs are 99% on the Mac, planning/testing/verification/manual coding/graphic design on Windows, committing and pushing to a repo both machines have checked out)
I haven't yet found a scenario where many Claudes and many Codexes running simultaneously on 35 concurrent features makes any sense, but I'd definitely encourage people to try multi-model cooperation since they all seem to have different sensibilities. I haven't made much use of Gemini in this context though because two's company, three's a crowd. YMMV.
I’m on the $100 Claude and $20 GPT plans. I almost never run out of weekly usage on Claude and occasionally blow my Codex allowance, but OpenAI lets you buy credits ad-hoc (either $40 per 1000, or just turn on and monitor the “top-off” and set it to buy like $5 at a time. The one or two times I’ve run out, by the time the week resets, I’ve only spent an extra 5 or 10 bucks.
Yes if you operate with worktrees, its actually possible to operate up to 5-10 at least I've succeeded with that multiple times.
I think whats important is, that you keep atomical small tasks and increments, and whenever possible merge things. to many hanging worktrees can quickly also become a nightmare managing
Leave it to the agent tbh, I spend more time on testing the worktrees - but even agents can do that for you - or if you add playwright MCP, make TDD both on frontend/backend and e2e pipelines - merge when cases work and tests approved
Superpowers + worktrees works really well for me. Superpowers works well at building comprehensive plans for implementing large features or refactors, and asks the questions up front, and then worktrees provide a safe place to actually perform that work.
It's not perfect; I've had some issues with Claude Code forgetting where it did things ("oh... it's not working because I'm not in the right directory"). I think it needs some architectural tweaks to function more reliably.
I expected this to become less necessary over time as models got faster, but the opposite has happened. It feels like Claude has actually gotten slower (but in fairness does more per prompt), meaning worktrees are even more essential now.
Depends on whether you can keep things separated logically. I have 3 git worktrees open, each working on a different area.
Generally its feature a, feature b and a refactoring branch of some kind.
My workflow is:
1. Add ticket in gitlab describing bug or feature in as much detail as possible along with acceptance criteria like expected unit tests or browser based tests.
2. In a work tree create a branch based on the id of that ticket in gitlab.
3. Start Claude, tell it to use a skill to pull that ticket, research and make a plan.
4. Review the plan, ask questions, refine.
5. Approve plan and let cluade cook.
6. Have Claude run a set of linters/tests/code quality checks and ground until done.
7. Start a new Claude instance, ask it to review changes made. Provide feedback to first Claude instance for changes.
8. Commit and push, creating a draft mr/pr in gitlab.
9. Review the actual code changes myself using gitlab. Comment on things not right.
10. Get Claude to use another skill to pull comments and work to resolve them. Also feed back any CI failures.
11. Manually close comments and push again. Repeat until done and ready for co-worker review.
I can only keep 3 threads like this going at once. Sometimes it’s only 1 or 2, depending on complexity. Smaller is better. Try to stay atomic and avoid feature creep in each mr.
Code can be logically separated, but my mind struggles to do the same. I guess this might require some training?
Pretty much everyone saying worktrees but I lean heavily on hot-reloading of both backend and frontend (to actually see what I'm doing) so it's too annoying to deal with dependencies and ports when things are isolated. Instead I keep everything on the main branch and just make sure to keep tasks pretty separate in scope, e.g. add this API route vs. fix this layout issue, so they don't step on each other. When I consider something done I have the agent commit what it worked on and start a new one.
I do use worktrees occasionally (especially during times where I'll have a very sticky problem that I make the LLM run in a loop on until it satisfies acceptance criteria, and want to isolate the potential fallout of Claudes Gone Wild), and I run Claude and Codex side by side, but I rarely have them work on truly-different tasks simultaneously.
The main reason is because if there's a significant bug or large optimization going on, that shit needs to be done, tested and merged before building more stuff on top, otherwise you run the risk of wasted time, tokens and effort having a bunch of parallel work running that may not end up compatible at the end.
Lately I've had a lot more success having Claude generate a plan, send the plan to Codex for co-validation/amendments, have Claude implement the plan, then have Codex PR review the commit (and likely make some edits of its own), then I test out the code/changes.
Meanwhile, my actual management of what I'm asking them to do is just a text file in Notepad where I'll write like BUG: xyz thing does abc or IDEA: let's change this to that as I'm testing in-app, with the actual code opened in Notepad++ tabs (lol feel free to roast me, I'm in front of 2 screens, one Windows (primary), one Mac (to the right), sharing keyboard and mouse -- LLMs are 99% on the Mac, planning/testing/verification/manual coding/graphic design on Windows, committing and pushing to a repo both machines have checked out)
I haven't yet found a scenario where many Claudes and many Codexes running simultaneously on 35 concurrent features makes any sense, but I'd definitely encourage people to try multi-model cooperation since they all seem to have different sensibilities. I haven't made much use of Gemini in this context though because two's company, three's a crowd. YMMV.
Using different models to supervise each other sounds reasonable. I’m curious what plans are you subscribed to for Claude and GPT?
I’m on the $100 Claude and $20 GPT plans. I almost never run out of weekly usage on Claude and occasionally blow my Codex allowance, but OpenAI lets you buy credits ad-hoc (either $40 per 1000, or just turn on and monitor the “top-off” and set it to buy like $5 at a time. The one or two times I’ve run out, by the time the week resets, I’ve only spent an extra 5 or 10 bucks.
Yes if you operate with worktrees, its actually possible to operate up to 5-10 at least I've succeeded with that multiple times.
I think whats important is, that you keep atomical small tasks and increments, and whenever possible merge things. to many hanging worktrees can quickly also become a nightmare managing
Do you manage worktrees manually or leave it to Agent?
Leave it to the agent tbh, I spend more time on testing the worktrees - but even agents can do that for you - or if you add playwright MCP, make TDD both on frontend/backend and e2e pipelines - merge when cases work and tests approved
Superpowers + worktrees works really well for me. Superpowers works well at building comprehensive plans for implementing large features or refactors, and asks the questions up front, and then worktrees provide a safe place to actually perform that work.
It's not perfect; I've had some issues with Claude Code forgetting where it did things ("oh... it's not working because I'm not in the right directory"). I think it needs some architectural tweaks to function more reliably.
I haven’t tried that, does it introduce some kind of “black magic” that makes the agent hard to observe?
It's just a set of skills and Claude commands. You can install via Claude plugins and read the prompts yourself
Yes, worktrees with workmux.
I expected this to become less necessary over time as models got faster, but the opposite has happened. It feels like Claude has actually gotten slower (but in fairness does more per prompt), meaning worktrees are even more essential now.
You can run multiple coding sessions technically, but human cognitive limits are the real bottleneck.
For someone running 5+ agents can you give an example of what each one would be doing when they are all running?