Today we are thrilled to announce INTHON v1.0, the production release of the agent-level programming language layer designed for AI tool orchestration and capability-bounded sandbox execution.

1. InthonVM: 63-Opcode Bytecode Stack Machine

In previous versions, INTHON executed source code via an Abstract Syntax Tree (AST) tree-walking interpreter. While excellent for debugging, loop-heavy tasks and large multi-step agent plans required higher throughput.

INTHON v1.0 introduces InthonVM, a custom virtual machine running an integer opcode stack instruction set across 63 canonical opcodes (Opcodes 0 to 62). Workloads running inside InthonVM execute up to 10–50x faster.

# Run script with high-speed InthonVM backend
inthon run --vm agent_plan.inth

# Compile to standalone binary bytecode container (.itc)
inthon compile agent_plan.inth -o agent_plan.itc
inthon run agent_plan.itc

2. Async Actor Mesh Scheduler

Modern agentic workflows require multiple specialized agents running concurrently. INTHON v1.0 introduces cooperative async actor scheduling via `inthon async-run`:

inthon async-run multi_agent_workflow.inth --timeout 300

The AgentScheduler manages asynchronous message passing, backpressure, and non-blocking event loops, enabling seamless multi-agent orchestration.

3. Interactive HTML Trace Replay (`inthon trace-view`)

Observability is essential for production AI agents. INTHON v1.0 introduces `inthon trace-view`, transforming JSON execution trace trees into interactive HTML dashboards with real-time timeline visualizers and cost breakdowns:

# Generate execution trace
inthon run --vm agent_plan.inth --trace-out trace.json

# Launch interactive trace dashboard
inthon trace-view trace.json -o dashboard.html

4. Summary of Improvements

Read InthonVM Specification →