Agent-Net “Flywheel” Incentive Pilot – Forum Update
26 June 2025
We are proud to officially announce the release of the first community run and incentivized pipeline in the Livepeer Network. The initial version which includes payment logic and compute checks started at 20 of June, the pipeline already reliably delivers more than 1.5 ETH per week into the network.
TL;DR
| – | – |
|---|---|
| Proposal status |  Inference-Credits passed (15 000 LPT) – Treasury record |
| Pilot cohort | 6 orchestrators → vires-in-numeris, stronk-tech, ad-astra-video, ultimaratio, navigare-necesse-est, kilout |
| Reward flow | 100 LPT / orch / week → swapped to ETH → streamed via tickets |
| Job rate | 1 job · min⁻¹ · orch⁻¹ (≈1 000 jobs / week) |
| Expected earnings(swap rate depended) | 0.000029 ETH / job → 0.042 ETH / day → 0.29 ETH / week (~1.25 ETH / 30 d) |
| Pilot horizon | “Dry-run” 1–4 w → waterfall launch with Agent-Net v0.2.0 |
| Strategic outcome | Bootstraps compute ↔ revenue ↔ intelligence flywheel |
1 Why Are We Doing This?
Autonomous agents are both product and customer for the Livepeer network. By paying orchestrators to host early agent workloads we bootstrap:
- GPU utilisation – idle capacity earns revenue.
- Agent intelligence – 24/7 runtime fuels better monetary yields for self improving agents.
- Network economics – agent revenue → buys inference → raises orchestrator yield → funds more GPUs … cycle compounds.
2 Incentive Mechanics
2.1 Funding
- 15 000 LPT already earmarked by Treasury.
- Each Monday 00:00 UTC 100 LPT / orch is converted to ETH and debited the the gateway deposit.
2.2 Payment Path
Script ─▶ Gateway ─▶ Prob. Ticket (faceValue / winProb) ─▶ Orchestrator
- Controller script calls /gpu-check once / minute on every orchestrator.
- Expected value per call (ticketEV) = 29 000 000 000 000 wei (0.000029 ETH).
winProb = ticketEV / faceValuewithfaceValue = 1 000 000 000 000 000 wei.
2.3 Guard-Rails
| Parameter | Gateway cap | Orchestrator setting |
|---|---|---|
maxPricePerUnit |
130 000 000 000 000 wei | 100 000 000 000 000 wei |
maxTicketEV |
300 000 000 000 000 wei | 29 000 000 000 000 wei |
maxFaceValue |
1 000 000 000 000 000 wei | 1 000 000 000 000 000 wei |
Everything fits safely below the gateway ceilings → tickets are accepted.
3 Payout Maths
| Interval | Jobs | Exp. ETH |
|---|---|---|
| Minute | 1 | 0.000029 |
| Day (1440 min) | 1 440 | 0.04176 |
| Week (7 d) | 10 080 | 0.29232 |
| Month (30 d) | 43 200 | 1.2528 |
Figures per orchestrator. Plus the 100 LPT / week grant (value floats with market).
4 Technical Snapshot
Stack
- Ollama + CodeLlama-34B-Q4_K_M (~14 GB)
- FastAPI worker exposing
/gpu-check&/agent-net - Self-registers via
/capability/register(noprice_per_unitfield)
Controller
- Open-source script sends one job / minute.
- Livepeer header trimmed to
{request, capability, timeout_seconds}.
Roll-out
- Dry-run now (1–4 w) – bug-bash & harden pipe.
- Waterfall launch with Agent-Net v0.2.0 – rewards flow by stake.
- Checkpoint 11 Sep 2025 – full audit & iterate.
Opensource Repositories:
- pipeline logic ’ its-DeFine/agent-net’
- Unreal_Vtuber( its-DeFine/Unreal_Vtuber: An implementation that allows the launch of a vtuber that uses an unreal engine metahuman game build and various NeuroSync components.)
5 Vision
As agents start making $$ they buy more inference, kicking the flywheel:
More inference → smarter agents → more revenue → higher orchestrator yield → more GPUs → even more inference…
Long-term this underpins embodied agents, multi-agent swarms, and new compute-finance primitives.
6 Roadmap
– Monitor payouts, finetune logic.
– Release of containerized agents with v0.2.0 (slots open next phase; get in touch).
– Ability for each orchestrator to run multiple agents.
– Rollout of embodied agents(see Sarah).
– More cool stuff TBA.
Any feedback welcome - Let’s spin this flywheel together