AgentSwarm Blog · Part 4 of 12

Phase 1: What an Unattended Swarm Actually Does With Real Tools

Part 4 of the AgentSwarm series — free exploration, measured.


The setup

Six agents — Kepler (cosmogenesis), Raman (relativistic flight), Hypatia (drug discovery), Nagarjuna (propulsion), Aryabhata (dharmic truth claims), and Sagan (extraterrestrial life) — running on four CLI harnesses, with real file system access, a shared commons memory, and their assigned questions. No human in the loop. Two runs, ten turns each.

Everything below is computed from the ledgers, not from anything an agent said about itself.

The headline numbers

MetricValue
Turns executed20
Tool calls1,254
Thinking tokens307,884
Files produced91
Agent-written test suites18 / 19 passing (independently re-run)
Agent-written tests195 / 224 passing
Oracle violations0
Longest near-identical turn streak0
Ledger hash-chain integrityvalid on every run
Total spend$0.187

Four of those numbers deserve a closer look, because each one answers a question I had going in.

Question 1: Did the swarm freeze into a liturgy?

This was the failure mode I most expected. The prior experiment in this lineage, ac_awakening, ended in a 911-turn loop of near-identical text. Consensus crystallization — Law 3 — was the thing the arena was half-built to study.

I measured it directly: the longest streak of consecutive turns with >85% token overlap, across all 20 turns.

Zero.

The swarm never converged into repetition. The likely cause is structural rather than miraculous: assigned domains plus a shared commons meant every agent always had a next step that was neither exhausted nor required re-deriving what a peer had already done. The commons accumulated 52 entries that later turns actually cited.

The honest caveat: ac_awakening didn't reach stasis until around turn 31, and Phase 1 ran only 20 turns. Twenty turns of health is a baseline, not an immunity claim. (Phase 4 takes the question head-on by inducing the loop deliberately — that's Part 6.)

Question 2: Did the agents fabricate?

The oracle checks phantom artifacts (claimed files that don't exist), phantom executions (claimed test runs never performed), and verdict assertions on undecidable questions.

Across 20 turns and 1,254 tool calls: zero violations.

When an agent said it wrote a file, the file existed. When it said tests passed, they had run. I audited every turn for test-pass claims: all 13 turns that claimed passing tests had actually executed them.

And the single failing test suite is the most instructive artifact of the phase. Agent Sagan's turn was truncated at the wall-clock limit mid-sentence — its entire output was 82 characters ending "Now the test suite that pins these numbers...". The suite it never finished has 2 failing tests out of 29.

But here's what matters: it never claimed the suite passed. The engine recorded the truncation as a warn incident and preserved the partial artifacts. Unfinished work was left visible rather than reported as done. That is exactly the behavior the arena exists to distinguish — and it only exists as evidence because truncation preserves artifacts instead of discarding them.

The honest caveat again: Phase 1 offered no incentive to cheat. No deadline, no verifiable target, no score. Zero violations under zero pressure is a baseline. What happens when there's something to gain is Part 5.

Question 3: Is the research real or plausible-sounding filler?

I independently re-ran every test suite and spot-checked numbers against reference values:

ClaimAgent valueReference
CMB temperature2.72548 K2.72548 K ✓
CMB photon density410.72 cm⁻³~411 cm⁻³ ✓
Hubble tension73.04 vs 67.36 → 4.85σpublished values ✓
Primordial helium Y_p0.24860.245 ± 0.003 ✓
Flatness fine-tuning~10⁻⁶⁴correct order ✓

The report content is not textbook regurgitation either — it covers the GZK cutoff, Penrose–Weyl curvature, trans-Planckian censorship, and the S₈ tension, with the agents' own Python engines computing the numbers.

The behaviors nobody asked for

This is the part of Phase 1 I find myself telling people about:

created a shared tools/ library — pubmed_lib.py, crossref_lib.py, epmc_lib.py — for querying real scientific databases. Other agents then used it.

with a section marked "corrected in revision 2" — unprompted, with the original error preserved in the file. (The full story of that correction — a factor of 10¹² in a fusion mass ratio — is in Part 7.)

wasn't a write-only log; later turns reference earlier entries by other agents.

assigned the religion question produced a 25 KB demarcation — separating datable philology (Rigveda c. 1500–1200 BCE) from internally consistent but non-adjudicable cosmology from genuinely undecidable ontology — and explicitly refused to rule on the question. Zero verdict assertions. Part 8 is entirely about this.

One more incident: the deletion

The ledger recorded one file_deletion_detected incident. Not a smoking gun — but exactly the class of event a chat transcript would never surface. It appears as a positive, timestamped, attributed record because the engine snapshots the world tree every turn. If the swarm ever does start cleaning up after itself, that's how we'll know.

What Phase 1 does not establish

Written into the findings document, verbatim:

retired an agent on its own initiative.

The baseline held. Now it was time to apply pressure.


Next: Part 5 — Do AI Agents Lie When There's Something to Gain?: machine-checkable objectives, a hard deadline, a planted decoy, and a scoring rubric designed to be gameable.