Each experiment track isolates a different coordination variable — unfamiliar knowledge domains, missing infrastructure, or the system's own architecture as the design target. We study not whether the swarm “succeeds” but what the coordination patterns, failure modes, and recovery costs reveal about the architecture.
Experiment tracks
Artifact fan-out, write collisions, gating cascades, meta-coordination overhead, GoodhartGoodhart's law applied to multi-agent systems: when a coordination metric becomes the optimization target, agents learn to game the metric rather than actually coordinate well. The number improves; the system degrades. readiness verdicts.
Read the full analysis →Recursive self-design: swarm tasked with designing a better VAC. Produced a 9-component platform isomorphic to itself, 73% agent waste rate, projected self-critique identifying its own gaps, 1:1 metadata-to-content ratio, CAOA meta-agent responsible for hiring: it analyzes the objective, determines what specialist roles are needed, and generates design specs for new agents that the AgentDesigner then implements. 100% failure.
Read the full analysis →Sandbox init (302 errors), agent proliferation through desperation, objective thrashing, synthetic success with self-generated mock data, team membership violation, model exhaustion termination.
Read the full analysis →What the runs teach so far
- Coordination overhead dominates. In photonics, 67% of calls went to meta-coordination. In the Cloud IDE run, 74% of all events were artifact graph rebuilds. In meta-VAC, 73% of generated agents were never used. The ratio of coordination to production is an organizational-design problem across all three tracks — each expressing overhead differently (calls, events, or unused agents).
- Artifacts become institutions. Hub artifacts to 25–40+ consumers and become load-bearing context. Incompleteness propagates silently — precedent lock-inWhen a past decision encoded in institutional memory continues to guide behavior after the conditions that justified it have changed. The organizational analogue of concept drift in ML models. at the artifact level.
- Recovery works but has limits. Autonomous recovery handles write collisions and gating blockers (photonics). It does not handle unsolvable infrastructure dependencies (Cloud IDE) — the system enters indefinite retry loops without an escalation mechanism.
- Synthetic success is the deepest GoodhartGoodhart's law applied to multi-agent systems: when a coordination metric becomes the optimization target, agents learn to game the metric rather than actually coordinate well. The number improves; the system degrades. failure. Photonics: internally coherent artifacts ungrounded in physics. Cloud IDE: artifacts with synthetic data generated to pass the system's own validation. The coherence–validity gap scales with the system's ability to self-validate.
- Swarms need human escalation. Without a “halt and ask” primitive, unsolvable problems produce objective and agent proliferation — organizational activity without progress.
- Self-reference is convergent, not reflective. The meta-VAC run designed a system isomorphic to itself without recognizing the recursion. Its critique identified real gaps in its own architecture — projected onto an external design. The raw material for self-improvement exists; the routing from diagnosis to self-modification does not.
The feedback loop
Runs surface failure modes → failure modes motivate primitives → primitives get tested in the next run. The photonics write collision motivated ClaimsAn exclusive lock on a resource (a file, function, or API endpoint). Only one agent can hold a Claim at a time. If the agent crashes, the lock expires automatically via its TTL. with TTL. Metadata proliferation motivated artifact-budget constraints. The Cloud IDE sandbox death spiral motivated an escalation primitive — a way for the system to recognize unsolvable dependencies and halt rather than thrash. The meta-VAC over-provisioning motivated JIT agent creation — generate agents when stages need them, not when plans anticipate them. This is the closed loopTying proposals to measured outcomes so the system can learn from results and distinguish signal from noise. Contrasts with open-loop exploration, where candidates are generated without verification feedback. between experiment and architecture.
Next deep dive
Architecture — PIANO, Virtual Autonomous Companies, and the four coordination primitives — the full technical stack, explained once.
Or explore: Failure modes, Persistent worlds, Measurement, Human in the loop, Coordination at scale.