Sāṅkhya described multi-agent systems three thousand years ago.

Indian epistemology has a tightly defined word for ‘valid means of knowledge’: pramāṇa. Different schools differ on how many pramāṇas there are — the Nyāya school catalogues four, Mīmāṃsā six — but they all agree on the underlying engineering shape. A claim is not knowledge until it is stamped by a process whose own correctness conditions are themselves articulated. Knowledge has a chain of custody.

An AI system that produces a confident answer without that chain has not produced knowledge. It has produced an output. The distinction is unfashionable to make in 2026 because it takes longer to build the chain than to ship the output, but it is the only distinction that ultimately matters once these systems are deciding things people care about.

Sāṅkhya enumerates twenty-five tattvas — categories — describing the resolution of awareness into a perceiving agent acting through perceptual instruments on a substrate. The tradition is usually translated in spiritual terms and then put on a shelf.

For the engineering reader, what matters is that the tattvas are an ontology of agentic computation: a separation between the unconditioned witness, a contextual self-model, an attentional bus, the discrete senses, and the substrate that furnishes the world they perceive.

That is a multi-agent architecture diagram. Buddhi is the discriminative function. Ahaṅkāra is the self-attribution module. Manas is the fast attentional router. The senses are the perceptual instruments. Puruṣa — the witness — is the role we are still failing to define in our own systems and which a great many AI safety problems reduce to misidentifying.

Two things, and both worth keeping. First, the Western tradition has built better tools for the scale of the verification surface — property-based testing, formal model-checking, statistical sampling at production volumes. Indian epistemology stops being prescriptive about method once the architecture is laid down; the West takes the architecture for granted and obsesses over the method. The two are complementary; the mistake is treating either as sufficient.

Second, the Western tradition is willing to be wrong out loud. Indian epistemology is patient and elegant; it is also, in places, unfalsifiable. What rescues the older corpus for engineering use is exactly the discipline that the modern stack imposes on it — does the contract hold under load? An assertion that cannot be checked is not, by Sāṅkhya’s own standard, a pramāṇa; it’s a wish.

The reason I am writing this and not, say, a paper on the loss function of a new agent benchmark is that I think we are five years away from a reliability crisis in AI agents that we will not solve by ironing out benchmarks. The crisis will be epistemic, not statistical.

People will start asking, in the same tone they asked it of evidence law two centuries ago: how do you know? what is your reference? what stamp did this answer pass through?

When that question becomes load-bearing, the systems that will win are the ones whose internal vocabulary already permits an answer. That vocabulary already exists. It was developed by people whose language most engineers find foreign and whose religious framing most engineers find awkward. Both are surmountable problems.