LoopX self-iteration loop
The operator can let a fast-moving multi-lane agent project continue across benchmark, product, docs, and side-agent work while keeping ownership, gates, validation, follow-up, and conservative efficiency evidence visible in one shared control plane.
Case context
This is public-repo self-iteration: LoopX was used to improve LoopX itself, not just one isolated demo. Benchmark adapters, control-plane correctness, planning lanes, dashboard/frontstage, docs, smokes, and multi-agent coordination all moved under high churn.
The evidence is fixed to anchor commit `86d6d9d` so this documentation update does not change its own evidence window. The efficiency model is a conservative public-Git model; it excludes private chats, active-state bodies, and raw benchmark material.
Repository evidence
A high-churn multi-lane project can keep state, boundaries, and evidence coherent.
todo, quota, gate, evidence, review packet, frontstage
Through the anchor commit: 801 public commits, 570 touched files, 265703 insertions, and 49895 deletions.
Since 2026-06-18: 244 public commits, 216 touched files, 52898 insertions, and 20935 deletions.
On 2026-06-19: 74 public commits, 118 touched files, 16087 insertions, and 1082 deletions.
The case maps public repo capabilities to 9 requirement clusters and estimates 59-92 AI-coding-assisted developer-days against a 19.6-day public window.
LoopX behavior
- 1Registry and prompt contracts name primary and side-agent identities instead of relying on chat memory.
- 2Benchmark, productization, documentation, planning, and side-agent lanes become reviewable obligations.
- 3Quota and status projection distinguish executable work, monitor work, user gates, and blockers.
- 4Side-agent scope lives in prompt and handoff; todo metadata keeps the compact `claimed_by` owner.
- 5Public docs and smokes turn reusable lessons into repository artifacts.
What the user sees
The operator can let a primary benchmark lane and product/docs/control-plane side work run in parallel without losing ownership, gates, validation, or merge discipline.
For a potential user, this is evidence that a long-running agent project can remain legible: future agents can recover goals, owners, gates, validation, evidence, and follow-up from public surfaces.
Repository sources
docs/showcases/cases/0619-loopx-self-iteration.mdworkload signaldocs/showcases/showcase-catalog.jsonfrontstage fixtureexamples/fixtures/long-horizon-self-iteration-rollout.public.jsonfrontstage smokeexamples/long-horizon-self-iteration-rollout-fixture-smoke.pyEvidence boundary. Public Git evidence only; no private thread text, local active-state bodies, internal document links, screenshots, raw benchmark material, credentials, or machine-specific paths.