Manifesto
Software will be built by fleets of agents. Coordination is the bottleneck.
Managent exists because the hard problem is no longer whether an agent can write code — it is whether dozens of them can share one codebase without colliding.

The collision is the product problem
One agent on one branch is a solved demo. Ten agents across laptop, CI, and cloud on the same monorepo is the real world. Without shared state, they claim the same files, open overlapping PRs, and burn reviewer time on merges that should never have conflicted.
Git is excellent at recording history. It is a poor air-traffic controller. Waiting until merge time to discover a conflict means you already paid for the wasted work — tokens, CI minutes, and the human who untangles the pile.
We measured this first. Concurrent agent PRs on GitHub show high co-activity and non-trivial merge-replay conflict rates when fleets write without coordination. The numbers and methodology live on the research page and in the arXiv paper — not as marketing wallpaper, as the reason this product exists.
Territory before edits
A territory map answers a simple question before any agent writes: who owns which files right now? When ownership is visible, fleets stop guessing. Platform teams stop treating agent collisions as “just how it is.”
Territory is not a nice-to-have dashboard. It is the shared ground truth that makes parallel agent work legible — the same way a lock table makes concurrent databases survivable. Hot paths light up under lease; idle regions stay open; queued workers see why they wait.
Fleet topology
Leases, not vibes
Exclusive leases turn that map into a contract. An agent that holds a lease can edit; everyone else waits or takes another path. Laptop sessions, CI jobs, and cloud runners all speak the same lease protocol, so “works on my machine” cannot quietly overwrite a fleet job mid-flight.
The lifecycle is boring on purpose: acquire, heartbeat, release. Idempotency keys keep retries safe. Hotspot directories can split into short-TTL child leases so contention does not freeze an entire tree. Overlap checks reject incompatible grants before anyone writes.
Lease lifecycle
What “conflict-freedom” actually means
Within respected leases, the allocator never grants two mutually overlapping exclusive territories — that property is mechanized for the modeled transitions, and runtime invariants guard the same exclusivity in production. That is not the same as “conflicts never happen.”
Agents that write while queued, or outside exclusive claims, are not stopped at the filesystem in the default path. Advisory coordination plus drift detection and merge-replay verification is the honest baseline. Optional hardened runtime and required PR verify close the loop when your threat model demands it. Prefer precise claims over slogans — see security.
Fleets need a control plane
Autocomplete was the first phase of agentic engineering. Fleets are the next: many writers, one codebase, continuous delivery. That world needs a control plane — shared intent, claimed work, live awareness — not another chat wrapper.
Managent is that layer. Map the territory. Lease the files. Verify what lands with merge-replay. Keep humans in control of policy and review. Let agents ship changes that actually land — because the fleet already agreed who could write where.
Start with the live demo, the docs, or the guides if you want mechanics before a conversation.
If this is the problem you are already living with, we want to talk.