How this was built

661 historical polities were researched, written, and validated without a single entry reviewed by a human researcher. The content corpus, structured JSON files covering states and empires from 3000 BC to 1945, was produced in roughly one day of wall-clock time. The person directing the build was on holiday, working from a phone.

The work was done by a coordinated fleet of AI agents running under Claude, Anthropic’s Fable model. A single coordinating agent ran the whole project without pausing for over 48 hours: planning, dispatching, reviewing, committing, and deploying in one continuous session. It spawned more than 300 subagents along the way. Nine regional scouts enumerated the 661-entry catalogue. Sixty-five research waves then processed those entries in groups of six agents working in parallel, each agent receiving a self-contained brief, writing a structured file, and exiting. A validation gate ran after every wave; all 65 waves passed at zero errors. No wave was retried. No batch was manually corrected by a human reviewer.

The adversarial review extended to the interface. A Playwright agent screenshot-audited the deployed UI against a written brief and blocked deployment until eight accessibility and layout findings were fixed and re-verified in a second pass. The go-live decision, meaning the merge to main and the production push, was made by a human. So were the catalogue approval and the editorial line decisions about what counted as a polity and how contested claims should be handled. Every other decision was automated or delegated to the coordinating agent under standing policy.

The Apex Spawn Loop

The core mechanism is the Apex Spawn Loop, a delegation pattern developed by Apex Intelligence: stateless waves of parallel agents, each owning a disjoint set of files, with a strict machine-checkable gate substituting for per-diff review and an auto-commit per wave. Statelessness is load-bearing. No agent knows what the previous wave’s agents did, so recovery from any single failure is a matter of re-running with corrected instructions rather than unwinding shared state.

The gate is the critical structural decision. An automated validator checked year encoding, source citation requirements, cross-reference reciprocity, basemap join policy, and a list of banned copy patterns. Agents whose output passed were committed immediately. Agents whose output failed were instructed to fix their own files and resubmit; the coordinator never patched content directly. That discard discipline kept the audit trail clean: every file in the corpus can be traced to a specific agent brief and wave commit.

The catalogue enumeration came first. Nine scouts assigned by world region identified the polities that warranted a full entry, and that scout phase is the one point where human editorial judgement was applied in advance of the research: the owner reviewed the combined 661-entry catalogue and approved it as the scope before a single research agent was dispatched.

Adversarial verification ran throughout the research phase. Each agent was briefed to flag contested claims with hedging language and inline source citations rather than resolve them with a single authoritative-sounding sentence. A separate class of fact-check subagents cross-referenced ruler dates and territorial claims against multiple sources. The label-lag adjudications, decisions about which polity claims a Cliopatria polygon whose label lags behind the historical succession, were made as coordinated editorial calls, documented in the session log, and enforced as standing rules for all subsequent waves.

Cost, speed, and the failure modes

The most significant cost in a corpus like this is not compute. It is the time a human editorial team would spend writing, fact-checking, and cross-referencing 661 entries to a sourced, consistent standard. That work would take months. The agentic approach produces it in a day of coordination at a compute cost several orders of magnitude below an equivalent human team.

What made it safe to run that fast was the audit trail built into the loop by construction. Every wave is a commit. Every adjudication, from source precedence to basemap join rules to the handling of contested boundaries, is a note in the session log. There is no ambiguity about what was decided, when, and why. That is not a documentation discipline imposed after the fact; it is a consequence of the workflow shape. A well-structured agentic loop produces its own audit trail.

The failure modes are worth naming directly, because they will recur in any deployment of this pattern:

  • Three agents went false-idle before writing their files. Each was resumed by the coordinator via a follow-up message restating the brief. Detection was manual: the coordinator noticed missing files in the wave summary.
  • One agent wrote its output to the wrong git tree, the main repository instead of the research worktree. Files were relocated. The wave gate then ran cleanly on the correct tree.
  • Duplicate cross-reference data slipped past the initial gate rules in earlier waves. The validator was updated mid-campaign; duplicate entries were caught on the next run and removed. No human reviewed the affected files individually.

All three failures were recoverable precisely because the loop is stateless and the gate is strict. Gated autonomy, meaning agents operating under machine-checkable rules with a hard commit gate, produced a more auditable output than unsupervised autonomy would have, and required less per-item human attention than manual review would have demanded. The trade-off is in designing the gate: the rules must be specific enough to catch real defects, not just format errors.

Apex Intelligence

World Empires was built by Apex Intelligence, an AI-native delivery and advisory firm. This site is one example of what that approach produces when applied to content and data at scale. The same patterns of Apex Spawn Loop coordination, gated autonomy, and adversarial verification apply to software delivery, research, and operational workflows.

If your organisation wants to work this way, apexintelligence.co.uk is where to start.