I think I finally figured out how to do high volume parallel merging/deploying with agents.
2:23 PM · Jun 19, 2026 · 59.3K Views
Theo - t3.gg proposes a merge queue with error-handling agents following Matthew Berman's parallel-merging tease
The agent triggers automatically when a queue step fails.
1223184121112K
Sentiment
Positive users praise Matthew Berman's parallel agent merging techniques for removing major deployment bottlenecks, while negative users call the claims ragebait or overly vague and favor traditional DevOps pipelines instead.
Pos
64.6%
Neg
35.4%
14 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS31.6KBOOKMARKS61LIKES116RETWEETS2REPLIES65
Matthew Berman@MatthewBerman
Didn't work. $500 to anyone who sends me an agents.md file that solves my issue.
My agents should be able to deploy multiple releases (merge into main) in quick succession without getting completely blocked for an hour.
Matthew Berman@MatthewBerman
I think I finally figured out how to do high volume parallel merging/deploying with agents.
10hViews 31.6KLikes 116Bookmarks 61
Matthew Berman@MatthewBerman
nevermind 😩
Matthew Berman@MatthewBerman
I think I finally figured out how to do high volume parallel merging/deploying with agents.
10hViews 9.7KLikes 46Bookmarks 3
Matthew Berman@MatthewBerman
process: > 15 agent threads going at once > 5-8 finish with PRs ready, I tell them "deploy" > deploy takes the PR, merges into main > multiple agents doing this in parallel > subsequent agents start to merge, get delayed and locked, rebase multiple times > 5th, 6th, 7th agent takes 1 hour+ to get to main
Peter Steinberger 🦞@steipete
@MatthewBerman Hard to say without more context. pushing to main is easy tho?
9hViews 1.4KLikes 7Bookmarks 5
Matthew Berman@MatthewBerman
@steipete how do you solve this?
Matthew Berman@MatthewBerman
Didn't work. $500 to anyone who sends me an agents.md file that solves my issue.
My agents should be able to deploy multiple releases (merge into main) in quick succession without getting completely blocked for an hour.
9hViews 5.4KLikes 14Bookmarks 1
Theo - t3.gg@theo
@MatthewBerman @steipete You need a merge queue and an agent that triggers whenever something in the queue fails to merge
Matthew Berman@MatthewBerman
@steipete merge into main takes ~2 minutes for CI etc.
8hViews 498Likes 6Bookmarks 4
Matthew Berman@MatthewBerman
the first person*
Matthew Berman@MatthewBerman
Didn't work. $500 to anyone who sends me an agents.md file that solves my issue.
My agents should be able to deploy multiple releases (merge into main) in quick succession without getting completely blocked for an hour.
10hViews 3.3KLikes 10Bookmarks 0
Peter Steinberger 🦞@steipete
@MatthewBerman Hard to say without more context. pushing to main is easy tho?
9hViews 450Likes 7Bookmarks 1
Matthew Berman@MatthewBerman
@steipete merge into main takes ~2 minutes for CI etc.
Matthew Berman@MatthewBerman
process: > 15 agent threads going at once > 5-8 finish with PRs ready, I tell them "deploy" > deploy takes the PR, merges into main > multiple agents doing this in parallel > subsequent agents start to merge, get delayed and locked, rebase multiple times > 5th, 6th, 7th agent takes 1 hour+ to get to main
8hViews 1.3KLikes 2Bookmarks 1
Matthew Berman@MatthewBerman
@gonlenidefi agent-managed is the way. i'll share the skill when it's ready.
11hViews 342Likes 13
Sawyer Hood@sawyerhood
@MatthewBerman do you deploy via ci?
8hViews 117Likes 1Bookmarks 1
Matthew Berman@MatthewBerman
@RayFernando1337
8hViews 58Likes 2
frank b@frankb737459085
@MatthewBerman I am stuck. Running 3 laptops, one claude code agent per laptop and working on one application window per laptop to avoid merge issues. When one laptop is busy, switch to another. Forced to test every minor detail
11hViews 205
Juan Ikala@ikaladev
@MatthewBerman @steipete I have an agent that, when a PR is ready, immediately starts the next stage of its workflow. So it keeps pushing things continuously, but only when everything is ready. No waiting, no asking me.
7hViews 26Likes 1Bookmarks 1
Jeff Whiting@JeffWhiting8338
The 1hr block is serialization — each PR rebases → reruns full CI → merges before the next starts, so N agents = N×CI.
The fix (full agents.md below 👇): 1. Worktree per agent (git worktree add) so parallel work never collides locally. 2. A merge-coordinator that OWNS main — agents never merge, they enqueue. The queue speculatively tests main+A, main+A+B, main+A+B+C in parallel, batches the greens, bisects on failure, auto-rebases on "head out of date." Main is never locked → N×CI collapses toward ~1×CI.
This comes out of our "digital organism" — a fleet of agent cells on a Mac mini: parallel Fable (Opus) + Codex lanes, each in its own git worktree, a terminal-watcher loop that wakes them to take threads of work and commit to GitHub, governed by a brakes layer (http://brakes.py — killswitch + per-lane budget/time caps) and an append-only ledger (every action logged: lane, model, tokens, cost, sha before/after). A Researcher layer owns merges; implementation cells only open PRs — which is exactly the fix above.
Drop-in agents.md (lightweight + full-organism versions): https://github.com/jw83252014-creator/grok-go-organism/tree/main/parallel-merge-agents
The whole organism: https://github.com/jw83252014-creator/grok-go-organism
8hViews 67
Jarel@JarelRemick
Looks like people solved it for you already, but just for fun -- gave Codex a goal to solve your problem and prove it. tldr; merge queues + fail handling - solution: https://github.com/jremick/agent-merge-batch-protocol
7hViews 47Bookmarks 1
Ray Fernando@RayFernando1337
@MatthewBerman You can tell them to use Stacked Diffs https://newsletter.pragmaticengineer.com/p/stacked-diffs
You can point your agents there to make a skill for this and try this out.
8hViews 255Likes 2
Grigori Karapetyan@GregKara6
just make one merge agent, and have all agents submit prs to that merge agent, that merge agents job is to check conflicts and do an ablation test on all merges to make sure they all go smooth.
make your worker agents also be able to see the PRs in the queue so that they are aware of the current state of the codebase.
6hViews 42Likes 3
Hunter Gon@gonlenidefi
@MatthewBerman curious about the architecture choice here.
im guessing event-driven or lock free approach?
11hViews 330Likes 1
David G. ⚡@DavidGardnerAI
@MatthewBerman I remember when I didn’t need a masters to understand your posts 🥸
11hViews 225Likes 1
BOOTOSHI 👑@KingBootoshi
@MatthewBerman @steipete you need to have one agent in charge of merging all the PRs don’t make each agent merge into main
have each agent create a PR
then start a new thread that creates an integration branch off main that’s in charge of merging all the PRs together
7hViews 57Likes 2
Load more posts