DISPATCHERxing-chiefSplits the stage. Merges P0 / P1 / P2. Does not edit code.Research3 AGENTS · READ ONLYxing-researcherFacts vs inference.xing-invest-complianceResearch, not advice.xing-security-privacySecrets and policy.look, check, reportProduct3 AGENTS · READ ONLYxing-ux-expertMobile, states, flow.xing-i18n-a11yEN / ZH / KO, keyboard.xing-search-visibilitySitemap, links.look, check, reportGrowth2 AGENTS · READ ONLYxing-growth-analyticsFunnels, experiments.xing-youtube-growthTopics back to sites.look, check, reportBuild3 AGENTS · MAY EDIT CODExing-feature-devConfirmed spec only.xing-bug-leadReproduce, then fix.xing-perf-leadMeasure one bottleneck.narrow write accessRelease3 AGENTS · GATEDxing-qa-leadTests, no code edits.xing-devopsGit only, after OK.xing-ops-monitorLive site vs HEAD.human OK before production
The structure: xing-chief dispatches, five groups report back. Specialists never assign work to each other.

The five groups

Fourteen specialists in five groups. Every card shows what that agent may touch — most can only look.

Research

3 agents

Facts before opinions.

  • xing-researcherRead-only

    Facts vs inference — labels which is which.

  • xing-invest-complianceRead-only

    Research, not advice.

  • xing-security-privacyRead-only

    Secrets and policy.

Product

3 agents

The site as the visitor meets it.

  • xing-ux-expertRead-only

    Mobile, states, flow.

  • xing-i18n-a11yRead-only

    EN / ZH / KO, keyboard and accessibility.

  • xing-search-visibilityRead-only

    Sitemap, links — how pages get found.

Growth

2 agents

Evidence from real usage.

  • xing-growth-analyticsRead-only

    Funnels, experiments.

  • xing-youtube-growthRead-only

    Topics back to sites.

Build

3 agents

The only group that writes code.

  • xing-feature-devCan edit code

    Confirmed spec only.

  • xing-bug-leadCan edit code

    Reproduce, then fix.

  • xing-perf-leadCan edit code

    Measure one bottleneck.

Release

3 agents

Nothing ships unverified.

  • xing-qa-leadRead-only

    Tests, no code edits.

  • xing-devopsGated

    Git only, after OK.

  • xing-ops-monitorRead-only

    Live site vs HEAD.

Permissions are the point

Most AI setups hand every agent every tool. Ours does the opposite — least privilege, enforced by structure rather than by trust.

10 of 15

agents are read-only

4

may touch the code repos

1

can run git — only after the founder says OK

  • Specialists report only to the dispatcher. They never assign work to each other.
  • The dispatcher splits the work and merges results into P0 / P1 / P2 — and never edits code itself.
  • QA runs tests but writes no code. Ops compares the live site against HEAD and changes nothing.
  • Releases go through xing-devops — git only — and only after a human OK.

Cross-checking: no single point of trust

The same 15-agent team runs in three independent coding systems — Claude, Cursor and Codex. If one system runs out of quota or goes down, another picks the work up and nothing stops.

On ordinary days the duplication is itself the test: the same task runs in more than one system and the results are compared. We pick tools by measured results, not by loyalty. The same rule runs underneath everything at XingAI — every key role has a second source, and every important number is checked against two.

One team, three systems — Claude · Cursor · Codex

  • Claude
  • Cursor
  • Codex

Questions, answered

How many AI agents build XingAI's websites, and how are they organized?
15 agents under one dispatcher, in five groups: Research, Product, Growth, Build and Release.
Which agents can change code?
Four of the 15 may touch the code repos. Ten of the 15 are read-only, and only one — xing-devops — can run git, and only after the founder says OK.
Who makes the final call before anything ships?
Humans set direction and make the final calls. Releases go through xing-devops — git only — and only after a human OK.
What happens if one coding system goes down?
The same 15-agent team runs in Claude, Cursor and Codex. If one system runs out of quota or goes down, another picks the work up and nothing stops.

When in doubt, ask XingAI.

A small human team. A standing agent team. Every step checked twice.

Meet the people →About XingAI →