@liveswarm haha all good, that makes sense now. lead owns the list, and if workers self serve they claim first. basically option 1 plus the shared board from my roundup
the real bug in multi-agent coding isn't merge conflicts.
it's two agents quietly doing the same task in separate clean branches.
worktrees fix files. you still need something that says "this one's taken."
how are you handing out tasks to your agents?
@lvgreenzhangyug good point, the lock can't catch duplicate features. i'd do both: a reviewer agent for the first pass, then a human signs off on anything touching shared modules
asked how you hand out tasks to parallel coding agents. best answers:
1. one "agent lead" per repo that spawns the rest
2. github issues + labels as the lock
3. one dispatcher assigning task ids
4. a shared board with clear ownership
worktrees isolate the output, not the intent. which one are you using?
@lvgreenzhangyug before. each worktree diff gets reviewed before it lands on main, so overlap shows up while it's still cheap to fix. feature module as the unit makes a lot of sense
@AlexChhk file claims shared across worktrees is a clean fix for exactly this. curious how it handles two agents that need the same file for different tasks
@lvgreenzhangyug yeah the silent overlap is the scary one. the lock works best when issues are scoped by area, not just files. then one review pass on the diffs before merge catches the rest
@zajca nobody really covered that, and it might be the hardest part. waiting isn't the same as done. i'd want the lead to check tests pass in the worktree before it marks anything closed
@surlymochan a judge that grades the same output differently twice isn't a judge, it's a coin flip. the robot arm number says it all, real tasks don't care about leaderboard rank
@clementrog one lead per repo is clean. you only ever talk to one agent and it handles who does what underneath. no task board needed when the lead is the board
@geminianmuskan depending what amount of money u have to consider . In india its not about the ideas anymore , anyone who has the MONEY to spend on infra and automations can do any business now a days.
@tiwarisuhani_11 recon it with offline files generated vai cronjob
use a separate api to read it instantly if instant results required, i see many more way