Sumerian Agent Patterns: Ancient Tablets, Modern Agent Design

What it is

Here's what got me going on this. About 4,000 years ago, Sumerian scribes were running something close to a multi-agent bureaucracy. Every transaction that mattered got sealed by a named person, dated to a named time period, reconciled on a schedule against a running total, and sometimes addressed and witnessed the way you'd address a letter. A seal, a timestamp, a periodic audit, an RPC-style header, a witness list: that's the same toolkit modern multi-agent AI frameworks are reinventing from scratch.

I didn't want this to be another post arguing the metaphor is cute. I wanted to test it. So I ran statistics against the SumTablets corpus, a transliteration dataset of 91,606 Sumerian tablets (6.97 million glyphs total), published by Simmons, Diehl Martinez, and Jurafsky in a 2024 paper for the ML4AL workshop at ACL. The question was simple: which of these ancient administrative patterns are actually real, statistically, and not just a story that sounds good? Whatever survived that test, I translated into named, single-responsibility software primitives, the kind with real inputs, outputs, and guardrails, that a modern agent runtime could actually use.

Let me make this concrete with one tablet. P101440 is a 39-glyph Ur III administrative tablet. It records a wool deficit and a day of female-labor accounting, gets closed with the seal of a named person (kišib₃ {d}šul-gi-i₃-li₂, "seal of Shulgi-ili"), and is dated by month and year-name. Strip that down to modern terms and you get a single write: a payload (the ledger lines), a seal (who wrote it), a witness list (empty on this particular tablet), and a resolved time period. Payload, seal, witnesses, period. That four-part shape is the sealed envelope the rest of this project is built around.

The findings

Over the course of this I catalogued something like 158 ideas. Nine of them earned the label first-class artifact, meaning they're backed by real evidence: statistical results with shuffled-baseline controls, cited tablet IDs, or a measured benchmark. Everything else in that catalogue is brainstorm-grade, useful for thinking out loud but not something I'd stake a claim on. The numbers below are the ones actually carrying weight.

Those nine findings map onto nine named agent primitives, among them TempleLedgerAgent, SealAuthorityAgent, AddressedMessageAgent, YearNameRegistryAgent, and RoyalDecreeAgent, and every one of them has cited tablet IDs backing its contract, not just a name that sounds good. I took one of the nine further than the rest: the sealed-envelope pattern behind kišib₃. I built it as a working, stdlib-only Python reference implementation, roughly 250 lines, and ran it against an anonymous-log baseline over 100,000 simulated writes. Here's what happened:

• Cost: about +37 tokens per write (that's an estimate), +7 microseconds of latency per write, and +59% more bytes at a 250-byte payload. That overhead isn't fixed, it shrinks fast: down to roughly 14% at 1 KB payloads and 1.4% at 10 KB payloads.
• Capability: the sealed log answered all 5 of 5 audit queries I threw at it (who wrote X, all writes by a given principal, all writes in a time period, integrity verification, and replay-after-cascade-revoke). The anonymous log answered 0 of 5. Not a partial gap. All of it.
• The hardest query: replay the log as of some point in time, after a parent principal has been revoked. The anonymous baseline just kept returning all 50 writes from that principal's descendants like nothing happened, a silent drift you'd never catch unless you went looking for it. The sealed log returned 0, correctly excluding every revoked descendant.

One more thing worth mentioning, because null results are honest and this one is useful: I ran a 99-skip by 5-genre by 1,000-shuffle equidistant-letter-sequence scan, the kind of test people use to go hunting for hidden codes in old texts, and it came back empty. Zero of 495 Bonferroni-corrected tests were significant. Keep that in your back pocket the next time somebody tries to sell you a "secret code in the clay tablets" story.

Why this might matter

I hold two framings of this at once, and I think both are fair.

The conservative read: this is a rigorous, reproducible corpus study, full stop. It applies real statistical controls (shuffled baselines, Bonferroni correction, a fixed random seed for reproducibility) to an actual ancient-language dataset and reports honest, falsifiable numbers, including one result that came back null. It also carries one finding all the way to a working reference implementation with measured cost and capability numbers against a real baseline. On its own, before you buy any of the bigger claims, that's a citable, auditable piece of empirical work.

The more ambitious read: the coordination problems these tablets were solving, attributing every write, naming every time period, reconciling accounts on a schedule, addressing messages to a specific recipient, are close to the exact problems modern multi-agent AI systems still get wrong: anonymous writes, flat memory with no structure, state that drifts silently. If a Bronze Age bureaucracy landed on the same shape independently, that's at least a data point suggesting these might be close to necessary primitives for any system, human or AI, that needs auditable multi-party coordination. A data point. Not proof.

I try hard not to overreach here. The observation that Sumerian administration looks like a proto-information system isn't new, it already shows up in popular essays, and I'm not claiming credit for noticing it. What I think this project actually adds is the empirical statistical mining behind it: cited tablet IDs, shuffled-baseline controls, numbers you can go check. Not the metaphor. The evidence.

What it isn't

• This is not a claim about "Sumerian thought." The corpus is more than 92% administrative records. Drawing conclusions about an entire civilization from this sample would be like judging a civilization by its tax receipts. Same problem.
• The sample is a slice, not the whole corpus. My statistical analysis covers a stratified sample of 2,069 tablets, 500 per genre across Administrative, Literary, Lexical, Royal Inscription, and Letter, plus 197 long-form tablets. That's about 2.3% of the full 91,606-tablet corpus. The findings are strongest for Ur III administrative tablets and Old Babylonian literary tablets, and weaker everywhere else. I'd rather tell you that than let you assume otherwise.
• This isn't a new philosophical insight either. People have been comparing ancient administration to modern information systems in popular writing for a while now. What I added is the statistical mining and the evidence trail, not the idea itself.
• It's not a cryptographic security model. Sumerian seals were physical objects, backed by witnesses and social standing, not math. The SealAuthorityAgent primitive borrows the shape of that system, named principals, revocation, witness sets, but not its threat model. Don't mistake a clay seal's guarantees for a cryptographic one.
• Year-names are not a neutral clock. They're political artifacts, named after royal acts, not a monotonic timestamp you can just trust.
• The lexical-list findings are the thinnest leg of this whole thing. They rest on only 69 tablets. The architectural slot for a taxonomy agent is real, but its actual content would need to come from external sources like CDLI or ePSD2, not from this corpus alone.

How it works, in plain English

Here's the method, in four stages. All of it is seeded and reproducible end to end (random_state=42, np.random.default_rng(42)), which means if you run it again yourself, you get the same tablets, the same templates, the same p-values.

1. Sample

I stratified-sampled 500 tablets per genre across five genres (Administrative, Literary, Lexical, Royal Inscription, Letter), plus up to 50 long-form tablets per genre. Total: 2,069 tablets plus 197 long-form.

2. Templates

For each genre, I measured structural-marker statistics, the physical-surface, column, row-ruling, and blank-space tags present in the transliteration. I pulled opening and closing line templates by position, surfaced the bigrams and trigrams distinctive to each genre using log-odds against the other genres, and ran hand-coded regex probes for known bureaucratic primitives: seal-of, year-formula, sum-total or audit, deficit, witness clause. Every template I extracted carries the tablet IDs it came from, so none of it is just my say-so.

3. Compression and equidistant-letter-sequence scan

I fit a per-genre Zipfian distribution and measured the compression-ratio delta between the raw token stream and a shuffled version of the same tablet. Then I ran an equidistant-letter-sequence decimation scan at skips 2 through 100, each one checked against 1,000 shuffled-baseline controls and Bonferroni-corrected across 495 total tests. That's the same family of test people use to go hunting for hidden codes in old texts, and I wanted an honest answer either way. Separately, I tested whether text on either side of a <RULING> marker shares more trigrams than a within-tablet shuffled baseline would predict, to find out whether it's a real logical row boundary or just a scribal flourish.

4. Mapping

For each empirical template that survived the statistical tests, I proposed a named, single-responsibility agent primitive with defined inputs, outputs, state, tools, and guardrails, and I labeled explicitly which parts are backed by data and which parts are speculative extrapolation on my part.

See the code and data

The repo splits its license by artifact type: the documentation and analysis outputs (outputs/*, the markdown reports) are CC BY 4.0, and the Python scripts (scripts/*, benchmarks/*) are MIT. The underlying corpus is SumTablets (Simmons, Diehl Martinez, and Jurafsky, 2024, ML4AL workshop, ACL 2024), also CC BY 4.0. Every tablet ID I cite in the findings, the P-numbers and Q-numbers, is a real CDLI catalog entry. You can look each one up yourself at cdli.mpiwg-berlin.mpg.de and aicuneiform.com.

I'd genuinely welcome cross-validation against CDLI, Oracc, and ePSD2, an extension to Akkadian or other periods, counter-examples to any template I've cited, or an independent reproduction of the Phase 3 statistics. Prove me wrong on any of it. That's the point of publishing the seeds and the scripts.

Where to go deeper

• Full code, data, and the reference implementation: github.com/NORTHTEKDevs/sumerian-agent-patterns
• The corpus and its paper: Simmons, Diehl Martinez, and Jurafsky, SumTablets: A Transliteration Dataset of Sumerian Tablets (ML4AL workshop, ACL 2024)
• Look up any tablet ID cited above yourself: cdli.mpiwg-berlin.mpg.de and aicuneiform.com

Built by Kristian Baer / Northtek. Analysis and documentation CC BY 4.0, scripts MIT licensed. Underlying corpus: SumTablets (Simmons, Diehl Martinez & Jurafsky, 2024, CC BY 4.0).