Skip to main content
Version: 0.2.16

Why X07

X07 is the deterministic, certifiable execution substrate for agent-written software.

The premise: as code generation gets cheap, the bottleneck moves to trust — deciding whether generated code is safe to run, and proving that decision to a reviewer. X07 is built around that bottleneck. Code that agents (or transpilers) produce runs sandboxed, budgeted, replayable, and provable.

The trust story (what makes agent code runnable and reviewable)​

  • Deterministic evaluation: keep correctness loops in solve-* worlds; use OS worlds only with explicit intent.
  • Record/replay: turn real OS interactions into deterministic cassettes you can re-run in fixture worlds.
  • Budgets: local budget.scope_v1 caps and arch-driven budget profiles prevent cost blowups from small agent edits.
  • Capability sandboxing: side effects are opt-in through explicit OS worlds and policy files; run-os-sandboxed defaults to a VM boundary on supported platforms.
  • Structured diagnostics: a 646-code diagnostic catalog with quickfix coverage enforced as a CI gate; the toolchain surface is JSON-first (diagnostics + patches + reports), so "lint → fix → re-run" is machine-drivable.
  • Spec-first testing: XTAL drives verify/repair/certify loops from pinned specs (XTAL).
  • Proof-backed certification: x07 verify produces proof and coverage artifacts, x07 prove check replays them, and x07 trust certify binds proof, test, boundary, and runtime evidence into a certificate (Formal verification & certification).
  • Review artifacts: semantic diff + trust report make changes auditable (world/capability deltas, budgets, nondeterminism flags).

Performance is part of the substrate story: X07 compiles via C to native code with fast compiles and small binaries, and a WASM target covers portable sandboxed execution.

The authoring story (honest status)​

Agents and humans can author X07 directly, and the 2026-06 toolchain improved that surface:

  • x07text: a lossless text projection (x07 ast to-text / x07 ast from-text; RFC 0001). Canonical source stays x07AST JSON.
  • behavioral summaries for stdlib exports in x07 doc, with fuzzy lookup
  • did-you-mean suggestions on unknown symbols, and structured diagnostics in x07 run failure reports

Direct authoring is an explicitly gated bet, not a settled claim. The comparative eval in labs/agent-eval/ — agents solving identical tasks in X07 vs Python/Rust — has a completed pilot (labs/agent-eval/results/pilot-2026-06-12.md) and a scaled protocol with a predeclared decision rule (labs/agent-eval/RUNBOOK.md). That run decides whether deeper language investment proceeds (RFC 0002: records, enums + match, string, f64) or X07 continues substrate-first. The project publishes the results either way.

Evidence pack (public, reproducible)​

1) Comparative agent eval (labs/agent-eval/)​

Agents solve identical bytes-in/bytes-out tasks in X07 and baseline languages, judged by the same vectors. The pilot result and the scaled runbook (with its predeclared go/park decision rule) are checked in.

2) Cross-language performance comparisons​

For runtime/compile time/binary size comparisons (X07 vs C vs Rust vs Go), use:

  • x07lang/x07-perf-compare (runs locally, verifies output equivalence)

3) Agent correctness benchmark harness (x07 bench)​

x07 bench evaluates patch submissions against versioned benchmark suites with deterministic artifacts and a machine-readable report:

  • Seed suite: labs/x07bench/suites/core_v1/ (expanded; recommended)
  • Docs: Benchmarks

4) Diagnostic catalog + quickfix coverage gate​

X07 tracks diagnostics as a catalog (646 codes) and enforces quickfix coverage as a CI gate:

  • Catalog tooling: x07 diag catalog, x07 diag check, x07 diag coverage
  • Rendered codes doc: Diagnostic codes

5) Agent-consumable spec + tool contracts​

The toolchain exposes stable, machine-readable surfaces for agents:

  • Schemas: spec/*.schema.json
  • Offline docs + agent portal endpoints: Agent contracts
  • Machine doc API: x07 doc --json ...

How to evaluate locally​

  1. Start with Install and Your first project.
  2. If you are operating through a coding agent, add Agent quickstart and The agent workflow.
  3. Run an end-to-end workflow on one of the reference projects under docs/examples/.
  4. Run x07 bench validate / x07 bench eval on labs/x07bench/suites/core_v1/.
  5. Reproduce the agent-eval pilot with labs/agent-eval/runner.py (stdlib-only, offline).
  6. Run x07-perf-compare for cross-language perf + build size comparisons.