Handoff queued — agent-31 takes over the moment agent-23 releases41 held · 17 queued
Scroll to explore
Scale the fleet to 48 and lease packages/core for agent-07, exclusive.
managent — fleet session
Fleet session
Forty-eight agents, one codebase, no collisions.
One control plane: inspect the territory, scale the fleet from 11 agents to 48 across eleven harnesses, and watch it as aggregate rollups rather than a wall of per-agent lines.
The scheduler is the product. It preempts a low-priority lease so a high-priority one can land, admits a 17-deep queue under CI backpressure, gates work on semantic impact, and batches every branch through one merge-replay pass.
The codebase, partitioned before anyone writes a line.
1,284 files across 42 globs, resolved into 41 exclusive regions. Regions cannot overlap by construction, so no two of the 48 agents are ever assigned the same line of code.
Unleased ground stays open, queued claims march at the border, and a dozen live branches batch into one merge-replay pass — including the group whose conflict is caught and replayed rather than escaping to main.
agent-09packages/core/src/lease.tsclaude · waiting on agent-07pos 1eta 0m18s•••
agent-31packages/ui/src/table/**gemini · waiting on agent-23pos 1eta 0m27s•••
agent-44packages/ui/src/table/**openclaw · waiting on agent-23pos 2eta 1m04s•••
agent-16infra/terraform/queue.tfcursor · waiting on agent-33pos 1eta 0m12s•••
agent-26apps/web/app/(fleet)/**claude · waiting on agent-12pos 1eta 0m41s•••
Leases
Every file an agent touches, held under an exclusive lease.
The ledger is the whole contract. A lease names one agent, one harness and one glob, and nothing else can write inside it until the lease is released.
At 48 agents that means 41 live leases and 17 waiters, ordered per glob with a position and an estimate. Requests that collide queue instead of failing, and hand over the moment the holder commits.
Verified against main before the merge, not after.
Managent batches every leased branch into groups, replays each group onto the current head commit by commit, and re-runs only the checks the change actually touches.
Batch #4182 took 12 branches in 6 groups. Group 4 hit a collision, was requeued into group 5 and replayed in the right order — caught in the replay, not in the merge queue. Twelve PRs landed, nothing escaped.
Managent came out of a study of how coding agents actually behave on shared repositories. These are the findings the product is built on.
Concurrency is the norm
PR- and repository-level co-activity rises steadily as the temporal-overlap window widens: agents working the same repository at the same time is the common case, not the edge case.
Figure 1 · RQ1
Conflicts cluster by file category
Across 167 conflicting agentic PR pairs, git merge-tree conflict types concentrate in a small number of file categories — which is what makes territory partitioning tractable.
Figure 4 · Taxonomy
Vendors collide with each other
Cross-vendor concurrency pairings show conflicts are not a single-harness problem: Claude, Codex and Cursor sessions overlap on the same files as readily as two sessions of one tool.
Figure 2 · RQ6
Exclusive leases, not advisory locks
A lease is held per file glob for the duration of an agent's edit. Laptop, CI and cloud runners resolve against one ledger, so two agents never hold the same line.
Leases · Ledger quorum
Merge-replay before landing
Every candidate merge is replayed against the current head before it lands, so a clean three-way merge that still breaks callers is caught before it reaches main.
Verification · Replay
Semantic impact is opt-in
Call-graph impact analysis runs only where a team asks for it. The default path stays fast; the expensive check is reserved for the changes that warrant it.
