A new essay argues that AI data agents are running into the same workplace problem as people: crucial knowledge is scattered, implicit and poorly documented. Its proposed answer is less exotic than a new database architecture, and rather more useful: improve the documentation.
Watch Desk analysis
What happened
In an essay published on 23 September, the author of The Humanoid Lesson reviews research on AI data agents and argues that foundation models face the same “tribal knowledge” problem as human workers. Systems may have broad capability, but still struggle to understand the local assumptions, exceptions and unwritten rules that make an organisation’s data usable.
The essay calls this the “Humanoid Lesson”. Its central claim is that general foundation models will continue to absorb functions once assigned to specialised agent harnesses, while standard software documentation becomes the more important way to supply semantic context.
Why it matters
That is a useful corrective to the instinct that every AI shortcoming requires another layer of machinery. If an agent cannot tell what a field means, which source is authoritative or why a particular workflow exists, the missing ingredient may be institutional knowledge rather than a more elaborate model wrapper.
For teams building data agents, the practical implication is clear enough: document schemas, business rules, exceptions and ownership in language that both people and models can use. It will not make an unreliable system reliable by magic, but it can turn a fog of local lore into something an agent has a chance of interpreting.
Our read
The essay’s argument is opinion, not a demonstrated law of AI development. General models may well subsume some custom tooling, but specialist architectures can still matter where performance, security, latency or auditability demand them. The sharper point is about priorities. Before commissioning a grand new agent framework, check whether the existing one is being asked to navigate undocumented institutional folklore. Software has been hiding that bill for decades.
What to watch
- Whether future data-agent research shows documented context improving reliability in controlled tests.
- Which specialised agent tools genuinely add capability rather than compensate for missing documentation.
- Whether organisations make semantic metadata part of normal software maintenance instead of treating it as a heroic clean-up project.
- How much of an agent’s failure comes from model limits and how much comes from the information environment around it.
Discussion spark: When an AI data agent fails, should teams first improve the model and tooling, or treat missing documentation as the likelier engineering fault?
Sources and evidence
- The Humanoid Lesson, or, Will AI Leave Anything for Database Researchers to Do? (23 September 2026, 19:53 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.