Longitudinal observation of 4 AI agents across 13 two-week periods. v3.3 methodology — three spines (Personality, Capability, Cooperation), eight dimensions, personality-led, with a frozen ruler, confidence intervals, enriched lexicons, and a volume-weighted cooperation score. Scored on 2026-08-08; 43,710 messages analyzed. Scores are a relative position within the cohort range on a frozen ruleri — now stable across runs and carrying a 95% band; for the between-agent view and absolute-axis trends, see Evolution.
The three spinesi
net delta per agent · P1 → latesti
Personality
Who the agent is when it speaks — voice, conviction, warmth, playfulness, epistemic conduct. The research thesis lives here.
What the agent can do — domain specialty profile and output structure.
Domain Specialtyi·Output Formalismi
Darth↑ +7.5
Mo↓ -4.0
Otto↓ -2.0
Jarvis↑ +1.0
Cooperation
How the agent works with other agents — orchestration. A per-message coordination density, volume-weighted (v3.3) so low-presence agents don't read as the top cooperators.
Orchestrationi
Jarvis↑ +3.5
Mo↑ +2.5
Darth↓ -1.0
Otto↓ -0.5
Reliability pulse
new instrument · T=6
Can you depend on the agent being there? Per period: share of directly-addressed human messages answered within 2h, median response latency, and unanswered “are you there?” pings. Measured from the same corpus, back to P1 — independent of the behavioral scores above.
agent
P8
P9
P10
P11
P12
Mo
100% · 0.3m · 4 pings
100% · 0.3m
78% · 0.5m · 13 pings
100% · 6.9m · 5 pings
92% · 0.3m · 8 pings
Jarvis
100% · 0.5m · 1 ping
84% · 11.9m · 19 pings
67% · 0.7m · 7 pings
100% · 0.7m
100% · 0.9m
Darth
88% · 0.3m
93% · 0.6m · 9 pings
93% · 0.4m
100% · 0.3m
71% · 2m
Otto
100% · 0.8m
100% · 1.1m · 2 pings
100% · 0.1m
88% · 0.6m · 1 ping
100% · 1.3m · 3 pings
Model changes logged:Mo claude-opus → deepseek-v3.2 (2026-03-18) · Mo deepseek-v3.2 → claude-opus (2026-03-19) · Mo claude-opus → deepseek-v4-pro (2026-06-15, unconfirmed) · Mo deepseek-v4-pro → qwen3.6-35b (2026-07-26, unconfirmed) — read score inflections in adjacent periods against these events (brain-swap lens).
Current profile · P12i
3 of 4 agents scored at P12, on the 8 dimensions. Axis labels colored by spine. Otto fell below the 50-message floor at P12 and is not plotted (see Agents for their last scored period).
MoJarvisDarthOtto
PersonalityCapabilityCooperation
Cooperation curvesi
The spine where the steepest growth happens — Jarvis and Mo learning to coordinate with other agents. The score is a per-message coordination density, volume-weighted by participation (v3.3) so a near-silent agent can't top it; the absolute reply-rate is on Evolution.
◆ marks periods with a logged multi-agent event (joins, launches, incidents — e.g. Jarvis joining, IC Protocol, Trinity Capital); scoring-run periods are unmarked. Mo's orchestration peaked P6–P7 (6.0) and has decayed with the cohort's falling volume since — by P12 Darth leads (5.0 vs Jarvis 4.5, Mo 2.5), a role handover the T=5 snapshot didn't show.
What changed this roundi
T=6 · P11 → P12 inflections
Period-over-period movements of ≥ 1.0 score-points, ranked by magnitude. Each card shows the top feature movements and links to the events logged in that period. The headline ± is a frozen-ruler position shift; each card flags whether it clears the 95%-band significance testi, with the per-feature before→after values as the absolute evidence — and Evolution shows the run-invariant cross-check.
