AI Event Fabric: Final Research Report
Motivation and hypothesis
Enterprise work is durable, collaborative, and auditable. The research question was whether a Kafka-compatible event stream can serve as an execution fabric for human and AI participants without making an AI supervisor the owner of process state. The hypothesis was limited: independently replaceable participants can coordinate by consuming durable WorkEvent records and publishing causally derived records.
Architecture and model
flowchart LR
H[Human participant] --> B[(Redpanda event log)]
A[AI participant] <--> B
O[Observer / replay projection] <--> B
B --> R[Future lifecycle research]The event model preserves event ID, work ID, actor identity, type, ISO-8601 time, correlation/run IDs, causal parent, payload, and payload provenance. Participants depend on input/output contracts rather than one another. Observers derive projections and lineage from retained records.
Method
The repository retains execution evidence for Experiments 01–04, 06, 07, 09, and 10, plus recorded inconclusive result summaries for Experiments 05 and 08. Deterministic participants isolate transport and participant behavior. Earlier live-provider configuration failures are retained alongside successful bounded Azure Sweden/Luna runs. See the final test results and failure analysis for the evidence-level distinction between direct errors and root-cause hypotheses.
Results
| Experiment | Status | Observed finding |
|---|---|---|
| 01 Multiple AI | PASS (prior evidence) | Independent proposals can react to one work event. |
| 02 Human review | PASS (prior evidence) | A human can participate through the same event boundary. |
| 03 Event chain | PASS | Fresh OCI-broker run recorded a four-event causal chain. |
| 04 Replay | PASS | Fresh replay projection matches all recorded event IDs and state. |
| 05 Failure recovery | INCONCLUSIVE | Recorded result summary reports two work events but one proposal; a readiness/assignment race was exposed. |
| 06 Live LLM | PASS | One real-broker Luna participant emitted a causally linked proposal. |
| 07 Hot swap | PASS | Deterministic and Luna participants emitted proposals at one contract. |
| 08 Competing consumers | INCONCLUSIVE | Recorded result summary reports empty worker logs and no performance values. |
| 09 Audit | PASS | Fresh OCI-broker lineage produced a Mermaid audit graph. |
| 10 SDLC topology | PASS | Fresh OCI-broker ten-event run ended in documentation. |
Operational boundary and future research
The experiments support an event fabric as a foundation for event transport, decoupled participants, causal history, and replay. They do not support a claim that an event broker alone is an AI execution runtime.
flowchart TB
F[AI Event Fabric\ntransport, contracts, causal history, replay] --> B[Architectural boundary revealed by experiments]
B --> R[Future research\nparticipant lifecycle, readiness, recovery, scheduling, observability]Experiments 05 and 08 make the boundary concrete. Durable history did not guarantee a worker was ready to consume it at the intended time, and the competing-consumer run produced no defensible measurement. A broker provides transport and partitioning, while stronger operational guarantees require future research into worker registration, lifecycle management, health monitoring, assignment awareness, recovery coordination, scheduling, and execution accounting.
Limitations and future research
The prototype has simple JSON validation, no schema registry, no authentication/encryption, no idempotency store, and no measured competing-consumer result. This checkout has one successful bounded live LLM result, but no multi-provider live substitution result. Future research should capture raw event artifacts, instrument group assignment, implement lifecycle experiments with explicit barriers, vary partition counts and hosts, collect latency distributions, and evaluate stronger lifecycle and operational guarantees.