Nexa v2.0.0 Released — Harness Native Runtime

After months of design and implementation, Nexa v2.0.0 is officially released. This is the most significant architectural evolution since the v1.0 AVM introduction — transforming the Harness six-tuple H=(E,T,C,S,L,V) from compile-time validation into runtime first-class primitives.

What Changed

The core design decision: the Harness is no longer just a compiler check. Each dimension now has a dedicated runtime subsystem that executes alongside your agent:

Dimension Syntax Runtime Tests
E autoloop HarnessKernel + AutoLoopConfig 52
T @tool ToolRegistry + ToolSchema 53
C with_context ContextManager + importance_weighted 52
S snapshot/restore StateStore + fork/merge 45
L before_step/reflect LifecycleHookManager 53
V verify EvaluationInterface + LLMRouter 59
Actor spawn/pass/await ActorSystem 18
WASM sandbox WASM Sandbox + full harness 15

Total: 296 new tests, plus v1.x 1500+ tests = 1800+ tests.

AVM Rust: 0 Errors, 0 Warnings

A major engineering effort went into fixing the AVM Rust backend. We resolved:

Result: cargo check0 errors, 0 warnings. Clean build.

12 v2.0 Examples

All examples compile successfully with nexa build --harness=warn:

examples/v2.0/
  01_autoloop.nx          — E-dimension: autonomous ReAct loop
  02_with_context.nx      — C-dimension: context management
  03_try_agent.nx         — E+L: fault-tolerant execution + reflection
  04_tool_annotation.nx   — T-dimension: zero-cost tool binding
  05_snapshot_restore.nx  — S-dimension: state snapshots
  06_fork_merge.nx        — S-dimension: branch exploration
  07_verify.nx            — V-dimension: output verification
  08_reflect.nx           — L-dimension: reflection injection
  09_lifecycle_hooks.nx   — L-dimension: lifecycle interception
  10_actor_system.nx      — Actor: multi-agent orchestration
  11_well_harnessed.nx    — full-dimension example
  12_harness_cli.nx       — Claude Code-like CLI framework

Nexa Code — The First Harness Native CLI Agent

Alongside v2.0, we're releasing Nexa Code — an interactive AI programming assistant built entirely on the Harness Runtime. It uses all six dimensions:

See our companion blog post: Introducing Nexa Code.

Backward Compatibility

v1.x code is fully backward compatible. v2.0 features are opt-in via --harness=warn:

// v1.x code still works unchanged
agent ChatBot { prompt: "..." }
flow main { result = ChatBot.run("hello") }

// v2.0: add Harness primitives incrementally
flow main {
    autoloop max_steps: 5 {
        with_context max_tokens: 50000 {
            try_agent {
                result = ChatBot.run("hello")
                verify result satisfies string
            } catch_correction(e: ToolError) {
                reflect "Adjusting approach"
            }
        }
    }
}

What's Next


— Owen, May 15, 2026