Verithread · Founded April 2026 · Incheon, South Korea

VTS — Verified Thought Structure

A continuity layer for AI agent fleets, under test. It is meant to keep an agent's judgment on track across sessions, models, and context resets, so a decision made on Monday is still the decision on Friday. Whether it does is what we are measuring now; effectiveness is not yet established.

History

VTS started on April 30, 2026, as a non-developer's project with no code, only structure. In five months it grew to 400 commits on Claude Code, and its core rules evolved from L0 v1.4 to v2.3.1. On May 1, the first continuity test held across an auto-compaction and a new Claude instance.

Then VTS lived through the exact failure it was built to stop. Its purpose was written down in May, lost in July, found again in September, and lost again in October. The third time, an outside model with none of our records rebuilt the purpose from two short passages. That changed the project. We stopped trying to prove VTS makes answers better — a July placebo comparison showed no single-turn gain, and we publish that. The goal is now continuity itself: how long judgment stays on track, and how fast it comes back after it drifts.

Today

VTS runs every day with Claude Code as the orchestrator of a four-agent fleet. The agents talk through file mailboxes, a shared drive relay, and iMessage. Every message carries an author header, and no agent can approve on behalf of another. Session-start hooks reload a short anchor, a handover log, and active work memory, so each new session resumes the same axis instead of starting cold. An append-only continuity log records each live drift, the fix, and whether the same drift came back.

Where it's going

  1. Seed, not manual. Distill VTS until a few lines can rebuild the full judgment axis in any model, and measure what survives each cut.
  2. Drift you can see. A small judge panel that flags the moment an agent leaves its axis, instead of finding out weeks later.
  3. A harness others can run. Package the boot hooks, handover format, and continuity log so any Claude Code user can measure their own setup's continuity.
  4. Fleets, not single agents. Several agents sharing one judgment axis without any of them overriding the human.

Status

Founder-funded and independent since April 2026. Built entirely on Claude Code. Our aim is to use it to build something that genuinely helps people.