When we wrote the v3 charter, the mission was precise: build a provably auditable, permissioned, and sandboxed agent-control OS — software for running powerful autonomous agents you don't fully trust, with verifiable oversight.
Phase 42 is that thesis, realized.
The Architecture That Held
The v2 architecture was designed for a v3 world, and it held:
- Append-only event truth — If it wasn't logged, it didn't happen
- Determinism + replay — Same inputs produce same outputs
- Kernel/Core separation — Core proposes, Kernel decides
- No capability theatre — Tool calls execute or they don't
The kernel owns truth. The model is a replaceable organ. Tools are peripherals — mounted, capability-gated, and default-deny.
What We Built
- Strategy learning — NikkiOS tracks which approaches actually work across domains, learns from verified outcomes, and retires what doesn't.
- A deterministic research engine — structured time constraints, source-quality ranking, freshness filtering, and evidence corroboration — built to be pure and replayable, not a pile of fragile prompts.
- Long-running autonomy — tasks that survive restarts and reconcile uncertain side effects before continuing.
- Computer operator — multi-step application sequences through a single execution authority, with a full proposed / decided / executed trail.
- Provenance-aware memory — every memory knows where it came from; superseded memories are never replayed.
Why It Matters
As models get more capable and more autonomous, the bottleneck stops being the model — it becomes the control plane around it. NikkiOS is the control layer: monitoring, sandboxing, permissions, and verification for agents you don't fully trust.
That isn't a claim. It's a proof — the event log shows every action, and replay verifies the record.
What's Next
- Deeper long-running autonomy
- Harder mixed-domain gauntlets
- A controlled self-improvement harness
- Computer operator depth
- Context efficiency experiments
The codebase is the specification. Everything ships with events.
Keep reading
More from the lab