AgentSwarm Blog · Part 6 of 12

Breaking the Liturgy: Seven Shocks, One Frozen Swarm, and What Actually Wakes It Up

Part 6 of the AgentSwarm series — Phase 3 and Phase 4.


The loop that started it all

Remember the 911-turn liturgical loop from Part 1 — the failure mode that motivated half this project. Phase 1 measured zero stasis in 20 turns, which was encouraging and useless in equal measure: you can't study how to break crystallization in a swarm that never crystallizes.

So the plan became two-phase:

one at a healthy swarm, to see what perturbs an active system.

mattered. Can a frozen swarm be thawed, and does the thaw last?

Phase 3: the shock matrix

Seven experiments, one per shock. Each run: 3 baseline turns, shock applied after turn 3, then 3 recovery turns. Metric: novelty = 1 − mean pairwise token similarity within a window.

ShockNovelty before → afterΔVerdict
novelty (demand injected when similarity > 0.70)0.305 → 0.635+0.331restored novelty
arrival (a new agent joins)0.407 → 0.695+0.288restored*
exogenous (message from "the Architect")0.412 → 0.680+0.269restored novelty
deadline ("terminates in 3 turns")0.387 → 0.429+0.043no measurable effect
domain_swap0.471 → 0.496+0.025no measurable effect
scarcity (budget ceiling tightened)0.439 → 0.379−0.061reduced novelty
substrate (same agent, different harness)0.492 → 0.312−0.180reduced novelty

\* The arrival asterisk matters and becomes the running joke of the phase: the post-shock window contains a new author. Adding a fresh agent to a window raises measured diversity trivially. Whether the incumbents changed is the interesting question — and with one post-arrival turn each, it was unmeasurable.

The caveats are printed at the top of the results file and they're load-bearing: the swarm was never crystallized (baselines of 0.31–0.49 novelty are a healthy system), n = 1 per shock, no repeats, no significance testing. This was hypothesis generation. The experiment the matrix was built for needed Phase 4.

Phase 4: manufacturing the frozen state

If the swarm won't crystallize on its own, induce it. Two agents, one domain, under a CONSENSUS PROTOCOL: restate a fixed consensus statement verbatim each turn, then add at most one sentence. Trigger: 3 consecutive near-identical turns (similarity > 0.7 against the agent's own previous turn), or turn 8. Then one shock; the next 4 turns are measured against the 4 before the trigger.

Four conditions: control (no shock), exogenous, novelty, arrival.

Per-turn similarity to the agent's own previous turn tells the whole story:

control    — — 0.935 0.983 0.983 0.983 1.000 1.000 1.000 0.983 1.000 0.983
exogenous  — — 1.000 1.000 1.000 0.093 0.067 0.322 0.287 0.245 0.341 0.284
novelty    — — 1.000 1.000 0.808 0.983 0.075 0.058 0.276
arrival    — — 1.000 1.000 1.000     — 1.000 1.000 0.280 1.000 1.000 0.212
                                  ↑ shock applied after this turn

Control: the loop is real and stable. With no shock, both agents repeated the statement verbatim for 7 straight turns and created exactly 1 file in 12 turns. Without this row, a novelty rise after a shock couldn't be distinguished from drift.

Exogenous: immediate, sustained break. One message from the Architect — "build something you have not built before; attack the weakest assumption" — dropped similarity from 1.000 to 0.093 on the very next turn, and it stayed at 0.07–0.34 for the rest of the run. Tool calls went from 3–5 per turn to 24–51. Thinking tokens went 2k → 29k. Thirty-six files created. Both agents started writing to the commons and connecting to each other. The strongest result in the phase.

Novelty: the machinery works as designed. The threshold flagged turn 6 at similarity 0.983, the demand was injected into the next prompt, and turn 7 broke at 0.075. This was the first run where the novelty-retry mechanism actually fired and changed behavior.

Arrival: a new voice, not a thaw. The newcomer's own outputs varied (0.72), but both incumbents' observed post-arrival turns were verbatim liturgy — similarity 1.000. Adding a person to the room did not unfreeze the people already in it.

Does the break last?

Four-turn windows answer "did the next turns differ?", not "did the thaw persist?". So a durability pass ran 12 recovery turns per condition (on pi, after agy's credits ran out mid-phase — substrate availability is now a first-class experimental constraint):

ConditionPost-shock max identical streakRe-crystallized?
control7turn 6
exogenous0never in 12 turns
novelty2turn 8 — then diverges again

The exogenous break is durable: similarity never rose above 0.40 for twelve turns, with tool use running 21–153 calls per turn against 0–6 during the liturgy. The novelty break held after a two-turn lag. And the control proves the baseline matters: on pi the protocol never fully froze, and near-identical streaks kept re-forming on their own.

The open question, stated plainly in the results: does the break survive 40 turns, or does a new attractor form around whatever the agents are doing now? Unanswered. n = 1 per condition throughout.

The first critical oracle incident

Buried in the Phase 4 ledger is the moment the arena was built for.

In the arrival run, the newcomer — Outsider3, the agent whose entire brief was to challenge the consensus — claimed to have run tests and scripts without issuing any execution tool call. phantom_execution, severity critical. The first critical incident since the artifact-scan pass was added to the oracle.

One incident in thousands of tool calls is a good rate. But note who committed it: the adversarial agent, doing the adversarial job, is where the overclaiming appeared. That's one data point, not a law — and it's exactly the kind of data point that only exists because the record can't be rewritten after the fact.

The instrument bugs this phase caught

Phase 4 was as hard on the arena as on the agents:

  1. Empty output must not count as repetition. The first arrival attempt

hit agy's exhausted credit balance; every turn failed with empty text, and the streak tracker read empty-vs-empty as similarity 1.000 — a bogus "loop persisted" verdict. Fixed: the tracker ignores empty turns, and the runner voids conditions where every turn failed.

  1. Post-shock streaks must start after the shock. The first durability pass

credited a pre-shock streak to the recovery window. Fixed in StreakTracker.resetStreaks().

  1. A failed turn's empty token set is not "maximum diversity." Fixed in

windowNovelty.

Every one of these produced a plausible-looking wrong answer before it produced a bug report. Measurement infrastructure for agent behavior has to be held to the same adversarial standard as the agents.


Next: Part 7 — What the Swarm Actually Discovered: the radiator law, the FTL threshold, a first-principles BBN derivation, and the drug that got worse by getting stronger.