Vats are Monte’s response to the vagaries of traditional operating-system-supported threads of control. Vats extend a modicum of parallelism and concurrency to Monte programs while removing the difficult data races and lock management that threads classically require.


From an entrypoint, the currentVat named argument will refer to the “top” or “first” vat:

> currentVat
Result: <vat(pa, immortal, 2 turns pending)>


This vat is named “pa”, is “immortal”, which means that it will never terminate computation abruptly, and has two turns of computation pending in its turn queue. All of this diagnostic information is Typhon-specific and may not be available in all implementations.

We can sprout a new vat at any time from an existing vat. The two vats will be distinct:

> def newVat := currentVat.sprout("re")
Result: <vat(re, immortal, 0 turns pending)>
> newVat == currentVat
Result: false

We can also seed a vat with a computation. The computation must be DeepFrozen, but otherwise any object can be used as a seed. This example is a bit dry but shows off the possibilities:

> newVat
Result: <vat(re, immortal, 0 turns pending)>
> def seed() as DeepFrozen { traceln("Seeding!"); return fn x { traceln(`I was sent $x`) } }
Result: <seed>
> def seeded := newVat.seed(seed)
TRACE: From vat re
 ~ "Seeding!"
Result: <promise>
> seeded<-(42)
Result: <promise>
TRACE: From vat re
 ~ "I was sent 42"
> seeded<-(object popsicle as DeepFrozen {})
Result: <promise>
TRACE: From vat re
 ~ "I was sent <popsicle>"
> seeded<-(object uncopyable {})
Result: <promise>
TRACE: From vat re
 ~ "I was sent <promise>"

Seeding produces a far reference to the result of the seed’s call, which might not be itself DeepFrozen. To interact with this reference, send messages to it. Note how sending popsicle caused the seeded object to receive a near (and thus printable) reference to it; this is because DeepFrozen objects travel between near vats directly.

What’s in a Vat?

The Browser Analogy

A vat, by analogy, is like a tab in a modern Web browser. It contains some objects, which may have near references between themselves, and a queue of pending messages to deliver to some of those objects. A browser tab might have some JavaScript to run; a vat might choose to take a turn, delivering a message to an object within the vat and letting the object pass any subsequent messages to its referents. Vats can be managed just like browser tabs, with vats being spawned and destroyed according to the whims of anybody with references to those vats. Indeed, vats can be managed just like any other object, and vats are correct with regards to capability security.

Vats, Formally and Informally

This is all confusing. What, precisely, is a vat?

Formally, a vat is just a container of objects. Vats have a turn queue, a list of messages yet to be delivered to objects within the vat, along with an optional resolver for each message. Vats compute by repeatedly delivering individual messages in the turn queue; each delivery is called a turn. Turns are taken in the order that they are enqueued, FIFO.

If a resolver is provided for a turn, then the resolver is resolved with the result of delivery. If delivery causes an exception, then the vat catches the exception, sealing it, and smashes the resolver with the exception instead. In either case, a membrane is applied to all objects which come into or leave the vat, including the result of delivery; this membrane replaces all non-DeepFrozen values with far references.

Informally, a vat isolates an object graph. Objects inside the vat can only refer to things outside the vat by far reference; there is no way to perform an immediate call across a vat boundary.

Whenever an object sends a message into a vat, the vat prepares to take a turn, whence the message will be delivered to the correct object inside the vat. Sends out of the vat produce promises for references to results of those sends, and the promises have normal error-handling behavior; if you send a message to another vat, and an exception happens in that other vat, then you’ll get a broken promise.

Vat Interface

Vats have two methods, .sprout/1 and .seed/1.

To sprout a new vat, call vat.sprout(name :Str) :Any, which returns a new vat. The new vat starts out empty, with an empty turn queue.

To put computation into a vat, call vat.seed(seed :DeepFrozen) :Vow, which does several things. First, the seeded vat copies the seed and its object graph into itself, isolating them from the calling vat. Then, the vat adds seed<-() to its turn queue, and returns a promise for that pending turn.


Vats are one of the more confusing parts of Monte, and some questions occur frequently.

So, no threads?

Correct. Monte does not have any way to block on I/O, so there is no need for threads at the application level.

Are vats parallel or concurrent?

It is implementation-dependent. Currently, Typhon is designed for an M:N threading model where up to M vats may take N turns in parallel on N distinct threads. However, Typhon currently only takes 1 turn in parallel. Other implementations may choose to do different parallelism models.

A key insight with vats is that a computation that is broken up into concurrent pieces on distinct vats can be transformed into parallel execution with maximal parallelism just by altering the underlying interpreter. The correctness of the computation does not change. This concept is from the actor model, which forms the theoretical basis for vats.

How do I perform parallel computations today?

Today, using Typhon, use the makeProcess entrypoint capability to run multiple processes to get node-level parallelism. We recognize that this is a very unsatisfactory solution for all involved, and we plan to eventually implement automatic parallel vats in Typhon.

For the future… Try to structure your code into modules; Typhon may parallelize module loading in the future. Also try to structure your code into vats, since we expect most interpreters to eventually implement parallel vat execution.

How do I perform concurrent operations?

Spawn more vats. All vats are concurrently turning. A vat will only ever lie fallow when it has no turns queued.

Why should we ever make synchronous calls?

In a nutshell, always make calls unless you intentionally want to create an asynchronous “edge” where your control flow stops, only to resume later. And also when you’re working with promises and far references, since you can’t make calls on those values!

Synchronous calls are very common. There are many kind of objects on which synchronous calls work, because they are near references. For example, all literals are near, and so is all operator syntax:

def lue := 6 * 7

There are many objects in the safe scope which are perfectly fine to use with either calls or sends.

Here are some handy idioms. To check whether a value is near:


A variant that might be more useful in the future:

value =~ n :Near

No, you misunderstood; why doesn’t Monte have only eventual sends?

Ah! There are several reasons, to be taken together as a measure of how difficult such a system would be to work with.

Some edges of Monte’s interaction with the external world are much better-modeled with calls than sends. A chauvanist argument can be made about how arithmetic should at least occasionally be lowered to a sequence of CPU instructions. However, we have found that a trickier and more important problem is dealing with object graph recursion, since Monte object graphs already can be quite treacherous. In Monte, object graphs can be cyclical and can hold delayed or eventual values. This poses a serious challenge, since sends for traversal can end up interleaved with sends which alter the structure or contents of the graph being traversed. Concretely:

  • Equality testing: x == y is a question that can, if they are Transparent, traverse the full transitive closures of both x and y.
  • Serialization: Pretty-printing, databases, RPC, DOT files, and all other serialization must traverse the full object graph as-is in order to not write out corrupted snapshots.
  • Hashing: Implementations may choose to define internal object hashes to speed up sets and maps. Application-level probabalistic data structures also often perform hashing. Like serialization, but just different enough to justify three sentences and a bullet point.
  • Garbage collection: GCs in the current state of the art are increasingly concurrent, running alongside mutators or only performing collections on per-mutator heaps. Nonetheless, when the GC would like to perform a collection, it often does need to traverse the object graph without worrying that an object will not race its own impending deletion with an incoming message delivery. This could be dealt with by requiring all sends to go through the vat turn queue, and pausing the vat in-between turns to collect. But then speed concerns pop up, and really this is a very deep rabbit hole…

So, for these reasons, we distinguish promises at the edges of our object graphs, and we implement these traversals using calls. As a practical consequence, uncalls are calls and must return near values. This also influenced the design of printers, which serialize by pretty-printing, and vats, which could optionally be implemented with per-vat GC.