Good agents, like good planners, keep options open. A coding agent tries three fixes for a bug. A travel agent holds four venue bookings until the client decides. An on-call agent weighs competing explanations for an outage while the evidence trickles in. The trouble is where those alternatives live. Usually in the prompt, which isn’t state at all, or in full copies like git worktrees and container snapshots, which have no idea which copy is more likely.
Multivergence puts the alternatives in the state layer. You fork the current state into cheap branches. Each branch has a weight, how likely it is; constraints, what must stay true; and a deadline, when it stops mattering. As evidence arrives, branches whose constraints break or whose deadlines pass drop away, and the survivor is promoted to the real state in one step.
The Pitch in Four Numbers
- About 4 KB for a branch that changes 3 keys of a 10,000-key state, against 2.1 MB for a full copy.
- 1,000 writer processes killed mid-change, 106 of them mid-promote, and each reopened to a whole state.
- 86.8 million reads checked against a model that copies every map, with 0 differences.
- 0 updates lost when 8 processes shared one store file 4,000 times.
How It Works
Six verbs cover it: fork, weigh, constrain, expire, observe and promote. Under the hood a copy-on-write map means a branch costs roughly what it changes, so holding dozens of alternatives is cheap. Every change is written as one checksummed record and synced before the call returns. Promotion is a single record, so a crash leaves the state either before the promotion or after it, never in between.
It comes as a Go library, in memory or backed by one file, safe across goroutines, standard library only. The multivergence command reads and changes the same file from the shell with JSON output, so agents written in any language can use it.
The demo shows an agent’s three travel plans collapsing to one as a fare search, a disruption feed and an expired seat hold arrive. The case studies cover a coding agent choosing between three bug fixes, an on-call agent narrowing an outage with weighted hypotheses, an agent holding four venue bookings and a dispatcher keeping four delivery plans open. Tests check isolation against full copies, kill writers mid-promote, replay deterministically, run a 30-day schedule of deadlines and constraints, and share one file between many processes.
Who It’s For
- Developers building agents that explore several plans, bookings or tool sequences before choosing.
- Planning and dispatch systems that hold competing schedules until reality picks one.
- Incident tooling that keeps several hypotheses alive with weights.
- Anyone simulating “what if” without copying the whole world each time.
Limits
It promotes one branch and discards the rest. It doesn’t merge branches. Branches are data only, so if a branch did something in the real world, like a payment, discarding the branch doesn’t undo the payment. It runs on one machine, with one process writing to the file at a time.
The Four-Question Screening
Every project featured here gets the same four questions, the way you’d screen a startup idea.
1. Where is the money in it?
Multivergence is open source under Apache 2.0, with no paid tier. The money is in the agent platforms being built right now. As agents take on longer, real-world tasks, they need a safe way to keep options open and commit cleanly, and platform builders pay for infrastructure that’s reliable. The paid paths are a hosted, multi-machine version for agent fleets, branch merging and audit trails for enterprise use, and support for teams embedding it. Booking, logistics and dispatch software are a second market that already thinks in alternatives.
2. What real problem does it solve?
Alternatives with nowhere to live. Keeping them in the prompt loses them. Copying the whole state for each one is expensive and says nothing about which is likelier. Committing too early throws away the better option. Multivergence makes holding and settling alternatives a normal operation on state.
3. The technology in plain words
Picture a choose-your-own-adventure book where you can keep a finger in several pages at once. Each finger has a note: how likely this path is, what has to stay true, and when it runs out. As the story reveals more, you pull out the fingers that no longer make sense, and when only one is left, that becomes the real story.
4. Moat and differentiation
Agent frameworks handle prompts and tools, databases handle state, and version control handles copies. None of them model weighted alternatives with constraints and deadlines that settle on evidence. Multivergence does exactly that, cheaply and crash-safely. It’s a new category, and the moat is getting there first with something that works and has the test record to prove it, then becoming the default state layer agent builders reach for.
Why It Matters
The more agents act in the world, the more they need to think in alternatives and commit carefully. Giving them a real place to hold those alternatives makes their decisions both better and easier to check.
Follow the Project
Read how forkable state lets agents try before they commit, start with the quick start, see how it was tested, browse the case studies, or get the code from the Multivergence repository on GitHub. Multivergence is listed with the other open-source projects on SoftDevStudio.com, and more featured projects are in the Projects section.