Ginnung · The observability cockpit
See how and why
your agents do what they do.
The observability cockpit for the Sonder runtime. Six faculties, one audit trail.
Compose the cognitive faculties your agents need. Watch the event stream. Answer the questions auditors actually ask. Ginnung is the user-facing control plane over Sonder — the runtime AI agents speak through.
The problem
Multi-agent systems fail silently. A 2025 analysis of 1,600+ execution traces across seven frameworks found that ~79% of failures came from coordination breakdowns, not reasoning errors. State-of-the-art systems hit only 25% baseline correctness.
Starting August 2, 2026, the EU AI Act requires high-risk systems to maintain logs sufficient to reconstruct decision context on regulatory demand. Fines run up to €15M or 3% of global annual turnover.
Every existing observability platform instruments the execution layer — timing, tokens, tool calls. None carry cognitive context: what the agent knew, what it was authorized to do, why it decided what it did.
What Ginnung is
In Norse cosmology, Ginnungagap is the primordial void — the substrate that exists before fire, before ice, before form. Everything else takes shape within it.
Ginnung is the cockpit for that substrate — the dashboard, faculty registry, and observability layer over the Sonder runtime that every faculty plugs into.
The architecture
Six faculties, bound by a shared event bus.
Sonder organizes agent cognition into six faculties. Every action emits a typed SonderEvent carrying structured context from all six simultaneously. Ginnung is the cockpit you watch them through.
| Faculty | Question | Product |
|---|---|---|
| Memory | What does it know? | Engram → |
| Reasoning | What is it concluding? | Parliament → |
| Governance | Was the action valid? | Lattice → |
| Capability | What can it do? | ACR → |
| Intent | What's the plan? | AWM → |
| Prediction | What outcome is expected? | Le-WM (coming) |
The audit log answers the five questions regulated industries require: what did the agent know, what was it authorized to do, why did it decide this, was the handoff valid, what did it predict? Read the SonderEvent spec →
The faculties, briefly
Memory
Engram
Persistent, queryable memory for AI agents. Vector + structured storage with confidence scoring and dream-cycle consolidation. Hosted at openengram.ai.
Open Engram →Reasoning
Parliament
Multi-agent deliberation engine with structured dissent. Eight reasoning topologies — debate, star chamber, jury, adversarial — with first-class disagreement.
Open Parliament →Governance
Lattice
Tiered circuit-breaker validation for agent handoffs. L1 structural contract checks, L2 semantic similarity, L3 judge confidence. Audit trails and shadow-mode rollouts.
Open Lattice →Capability
ACR
Agent Capability Runtime. Manages what tool instructions load into context, at what resolution, within what budget. The OS between an agent and its tools.
Open ACR →Intent
AWM
Adaptive Workflow Monitor. Captures step traces, signals regime changes, emits planned-action records before execution.
Open AWM →Prediction
Le-WMComing
JEPA-based world model for outcome prediction with Bayesian confidence. Latent-space planning at a fraction of foundation-model compute.
Who Ginnung is for
- Engineering teams building agents that touch financial trades (MiFID II, FINRA 2026), patient data (HIPAA), or anything regulated by the EU AI Act.
- Researchers needing reproducible cognitive context across multi-agent experiments.
- Anyone who has ever asked an LLM agent “why did you do that?” and gotten a worse answer than they gave.
Built by
Ginnung is built by heybeaux, a one-person studio in Powell River, BC. The stack is open-source, the spec is public, and the audit log is the product.