I've been building VVM: a language for agentic programs where the LLM is the runtime.

Most stacks today keep orchestration in Python/TS and treat the model like an API. VVM inverts this. You write a `.vvm` program, paste it into an agentic session with the spec, and the model simulates the VM and executes your program. This is configured via a plugin, no SDK or build step required.

```

agent researcher(model="sonnet", skills=["web-search"])

agent writer(model="opus")

papers = @researcher `Find recent papers on {topic}.`(topic)

# semantic predicate — semantic if condition

if ?`sufficient coverage`(papers):

  report = @writer `Synthesize findings.`(papers)

  papers = @researcher `Expand search.`(papers)

```

It's open source and current runs as a Claude code plugin and the repo includes 37 examples from basic agent calls to advanced patterns that showcase building resilient workflows.

Still pre-alpha -- the spec is fairly complete, but tooling is early. Mainly looking for feedback from people building agentic systems who've felt the impedance mismatch between imperative code and what agents actually need.

https://github.com/karanchawla/vvm