Consiliences AI is the system that runs everything else. The Observatory’s signal corpus, the Institute’s papers, the Angles & Footnotes newswire, the Phronopolis essays — none of them are produced by a person sitting down to produce them. They are the output of a multi-agent system I have been building for about five months.

This page is the honest account of what that system is, what it does, and what it is disciplined against.

§ I   The shape

Start with the shape. Forty persistent agents run continuously. They are not chatbots waiting for a prompt. Each one has a standing goal, a sandboxed working directory, and a defined trust level.

Many of them work the same core loop: they observe — RSS feeds, data APIs, the signal corpus, each other’s output — score what they observe against their goal, discard most of it, and surface or act on what survives. That observe-score-discard-or-surface loop is the heartbeat of the platform, and it runs thousands of times a day without anyone watching it happen.

Other agents do different work — running multi-step research plans, synthesising across domains, auditing the system itself — but the principle is constant: a standing goal, a sandbox, and a trust level that bounds what the agent may do unattended.

§ II   Scale and sovereignty

The platform is roughly 287,000 lines of Python excluding tests, with nearly six thousand test cases holding it in place. It maintains and progressively validates a corpus of more than fifteen hundred research signals against a falsification battery; over half of that corpus has not yet been battery-tested, and the kill list is published rather than hidden.

Language-model calls route across several cloud inference providers for performance, but the only mandatory endpoint is a local server I control. Cloud routing is never authoritative; the system degrades to local-only without losing correctness. That is a deliberate sovereignty decision, not an accident of cost — though the effect on cost is real: the platform has run on pennies a day throughout 2026.

§ III   Consumer and provider

Around three dozen external data and service integrations feed the agents — financial market data, climate and agricultural datasets, search, the ordinary infrastructure a research operation needs.

Three Model Context Protocol servers expose the platform’s own surfaces in the other direction: one publishes the Observatory’s validated signals, one bridges directly to the agent fleet’s research pipeline, and one provides the platform’s web-fetch tooling. The platform is both a consumer of data and a provider of it.

§ IV   What keeps it from being slop

What keeps this from being AI slop is not the intelligence of the models. It is the discipline around them.

Every agent operates inside a graduated trust gate — supervised, semi-automatic, monitored, autonomous — and promotion between tiers is a human decision, never an automatic one. Every signal the Observatory confirms has survived a falsification battery; about one in six candidate signals is killed outright, and the kill list is published rather than hidden. Every consequential action an agent takes is written to an append-only audit log.

The system was built to be safe and legible before it was built to be fast — and when those two goals conflict, legibility wins.

§ V   One engine, four arms

The same engine runs four public arms. The Consiliences Institute publishes long-cycle structural research. Angles & Footnotes runs a multi-perspective newswire. Phronopolis publishes essays from named historical personas. This subdomain is the engine itself.

The standard the operator applies to weak evidence in an Institute signal is the same standard the operator applies to a wobbly lens in an Angles & Footnotes story. The mechanism differs by arm; the standard does not. Consistency of standard across all four is the product.

§ VI   Where to go next

This subdomain documents the architecture in more depth: how the agent mesh works and why the trust gate is the load-bearing design rule, how the validation framework runs as code rather than as editorial taste, the track record in honest numbers, what the engine does beyond publishing, and how to engage. Each of those pages is linked from the navigation above; this overview is the place to start.