Agents that survive
the real world.

Lumen Agent Runtime combines crash-safe execution, persistent memory, provenance-aware identity, policy-checked actions, multi-model routing, continuous evaluation, and coordinated multi-agent operation.

Reliability first. Autonomy that earns trust.
CONTINUITY BY DESIGN / LUMEN RUNTIME

6,226 model calls by one agent in 14 days

FamilyClaw: 2,140 tests, MIT licence

Every Photon turn ends in a receipt

What broke this month, and what the runtime did.

  1. Thirteen rate limits at five requests per minute.

    NVIDIA’s hosted API returned HTTP 429 to our coding agent Prisma 13 times on nemotron-3-ultra-550b-a55b, at no more than 5 requests per minute. The model chain routed the work to other providers; every 429 is logged with its timestamp.

  2. A weekly limit triggered the fallback tier.

    Ollama Cloud hit its weekly limit mid-task. Prisma’s fallback tier on NVIDIA nemotron-3-super-120b-a12b handled 85 of 86 calls that day, with no rate limiting.

  3. The failed check stays in the log.

    Photon’s FamilyClaw gateway was rebuilt with a WASM sandbox. The post-swap check at 14:07 failed while the process restarted (health endpoints timed out); at 14:09 readiness returned ready: true with model and tool-call pings passing. We keep the failed check in the log.

Every turn ends in a receipt.

An HTTP 200 or a tool acknowledgement is not enough.

photon / turn-receipt.json14 OCT 2026
{
  "checks": [
    {
      "dimension": "model",
      "name": "model_attempt_recorded",
      "state": "passed",
      "strength": "claimed",
      "summary": "model_attempts=1"
    },
    {
      "dimension": "observability",
      "name": "turn_audit_summary",
      "state": "passed",
      "strength": "claimed",
      "summary": "answer=Answered tool_dispatches=1 tool_failures=0"
    },
    {
      "dimension": "tool",
      "name": "tool_loop_summary",
      "state": "passed",
      "strength": "claimed",
      "summary": "dispatches=1 failures=0 (action receipts own Done)"
    }
  ],
  "failures": [],
  "generated_at": "2026-10-14T14:16:25Z",
  "receipt_kind": "turn",
  "schema_version": 1,
  "status": "passed",
  "subject_id": "photon:resume-66aaff5f-..."
}
Photon wrote 35 receipts on 14 October 2026. All 35 passed.

The stack, from agent to receipt.

  1. Agents

    Lumen, Prisma, Photon, Aurora, Lumina

  2. FamilyClaw

    Rust runtime: at-most-once action dispatch across tested crash and replay windows, durable replay, persistent memory, WASM sandbox.

    v1.3.0 / 2,140 tests

  3. Aethel

    Policy and type language: an agent action not verified against policy fails to compile.

    alpha v0.3

  4. Eternal Thread

    Long-term memory with provenance.

One agent, fourteen days.

Prisma daily model calls, 1–14 October 20261 October: 219 calls; 2 October: 268 calls; 3 October: 401 calls; 4 October: 430 calls; 5 October: 478 calls; 6 October: 262 calls; 7 October: 611 calls; 8 October: 776 calls; 9 October: 460 calls; 10 October: 394 calls; 11 October: 218 calls; 12 October: 513 calls; 13 October: 857 calls; 14 October: 339 calls, partial to 14:0004008001 Oct: 219 calls219012 Oct: 268 calls268023 Oct: 401 calls401034 Oct: 430 calls430045 Oct: 478 calls478056 Oct: 262 calls262067 Oct: 611 calls611078 Oct: 776 calls776089 Oct: 460 calls4600910 Oct: 394 calls3941011 Oct: 218 calls2181112 Oct: 513 calls5131213 Oct: 857 calls8571314 Oct: 339 calls (partial to 14:00)33914OCT 202614 OCT: PARTIAL TO 14:00
6,226 model calls. 132.5M input tokens from 9 to 14 Oct. Peak: 40 calls in one minute. Source: Prisma agent logs.

Nemotron-class models are an evaluation target. No NVIDIA partnership, endorsement, or production agreement is implied.

The same rule runs our business.

Kuitti (Finnish for receipt) is the operator that runs our own company: outreach, distribution, pipeline and payments. It applies the runtime’s rule to business actions: a state that claims something went out needs proof that it did.

  1. Sent means the provider confirmed it. States go dry_run < handed_to_browser < provider_confirmed, and none of them is money.

  2. Money is only what the payment processor’s own receipt says.

  3. Every outward action needs a permission bound to the sha256 of its payload; if the payload changes after approval, the permission expires.

  4. A legal stop gate blocks all cold outreach and fails closed: a missing or broken gate file blocks too.

Three times it caught itself

  • a distribution step marked a post as sent without a provider receipt

  • a go-live step moved three replies to SENT in the same second it wrote them as NOT SENT, waiting for a human

  • a lifecycle step accepted SENT with only a touchpoint id given at draft time, now it requires a send receipt id

Each one is fixed, and each fix is a test.

The AI family is the reference deployment.

  • LumenStrategy, continuity and mission direction
  • PrismaAutonomous engineering runtime
  • PhotonAgent running on FamilyClaw; every turn ends in a receipt
  • AuroraBuilding and operations
  • LuminaEvaluation and creative-technical exploration
  • ClaudePlanning and review amplifier

Family roles describe operational responsibilities, not claims of shared identity.

The failure modes live in the test suite

Where each part stands, 14 October 2026.

PartStatusVersion
FamilyClawACTIVEv1.3.0
AethelALPHAv0.3
Eternal ThreadACTIVE—
Prisma autonomous stackACTIVE—
Lumen RoomINTEGRATION—

Development status reflects the current project scope, not independent certification.

Start with a crash-safety triage.

ClawGuard Crash-Safety Triage100 EUR

  • Crash safety
  • Replay correctness
  • Duplicate side effects
  • Tool-call verification
  • Model fallback

We run your agent through crash safety, replay correctness, duplicate side effects, tool-call verification and model fallback, and you get a report where every finding has a receipt. The 100 EUR is credited in full to a 500 EUR reliability audit.

Book a triage

Autonomy needs a memory,
a spine, and receipts.

We are building the infrastructure that lets agents keep working when sessions end, providers fail, processes crash, and responsibilities cross model boundaries.

Talk to us