VLM’s Orion 2 skips the standard agent playbook — instead of calling tools one at a time in a loop, it generates a full program and runs it end to end.
The pitch is straightforward:
- Single program generation instead of sequential tool calls — fewer round-trips between the model and execution environment
- Lower latency because the entire workflow executes as one unit, not a chain of API hops
- Composable, inspectable, deterministic workflows — the orchestration layer is code, so you can read it, version it, and replay it
Most agent frameworks today follow a ReAct-style loop: the model picks a tool, calls it, reads the result, picks the next tool, repeats. Each hop adds latency and a chance for the model to lose context. Orion 2 collapses that loop into a single code-generation step. The model writes the program; the runtime executes it. That’s a meaningful architectural difference — it trades flexibility for speed and predictability.
The “visual” part of the name refers to vision-language model capabilities. VLM (the company) focuses on multimodal AI that can process images and documents, so Orion 2’s code-mode approach applies to workflows involving visual inputs — think document extraction, image analysis, or UI understanding — not just text-based tool use.
You can try it at chat.vlm.run and watch a demo video showing the workflow in action.
NOTE
Orion 2 surfaced on Hacker News as a Show HN with limited discussion so far. No pricing details or model benchmarks have been published yet — treat this as an early-stage tool worth watching, not a production-ready replacement for existing agent frameworks.
Code-as-orchestration is the interesting bet here. If the generated programs are genuinely inspectable and deterministic, that solves one of the hardest trust problems in agent systems — knowing exactly what the agent did and why. Available now at chat.vlm.run.