Autonomous Multi-Agent Governance, Intelligent Failover & Pre-Debit Budget Guardrails on Mozilla Otari Gateway Built for the MozFest / Mozilla.ai "Developers Wilding: Online" Hackathon 2026 ($1,000 1st Place Track)

Autonomous multi-agent swarms consume massive token volumes across continuous execution loops. Operating agents against commercial frontier APIs introduces critical institutional risks:

Unbounded Wallet Drain: An agent caught in an infinite self-correction loop or receiving malicious prompt injections can rapidly exhaust thousands of dollars without spending controls.

Provider Lock-In & Outages: Upstream provider outages (e.g. HTTP 429 rate limits, 502/503 service disruptions) crash autonomous pipelines unless resilient failover cascades exist.

Overpaying on Low-Complexity Tasks: Routing trivial operational tasks (e.g. data scraping, classification, heartbeat checks) to $0.015/1k frontier models burns up to 90% of an organization's compute budget unnecessarily.

Otari Agent Sentinel is an open-source autonomous agent control plane built directly on Mozilla.ai's Otari LLM Gateway. It provides:

Adaptive Semantic Complexity Classifier: Automatically categorizes incoming agent tasks into 3 tiers (FAST_OPEN_WEIGHT, BALANCED_CODE, REASONING_FRONTIER), directing simple tasks to sub-millisecond open-weight models (Llama-3.1-8B, Mistral-7B) and reserving heavy reasoning (DeepSeek-R1, Claude-3.5-Sonnet) for formal verification and multi-step math.

Otari Pre-Debit Budget Governance: Implements Mozilla Otari's virtual API key reservation protocol. Before any LLM inference call is dispatched, the sentinel reserves an estimated token budget ceiling. If an agent exceeds its quota, execution is cleanly blocked before tokens are consumed.

Circuit-Breaker Failover Cascades: Continuously monitors endpoint latency and error rates. If an upstream model experiences an outage or 429 rate limit, requests automatically cascade to healthy alternate models within milliseconds without crashing the agent pipeline.

Real-Time Cost Savings & Telemetry: Computes verifiable token spend vs a standard frontier benchmark ($0.003 in / $0.015 out per 1k tokens), demonstrating 60% to 95% capital savings.

Loading

graph TD subgraph Autonomous Agent Mesh AgentA[DeFi Market Scanner] -->|Task Request| Sentinel{Otari Agent Sentinel} AgentB[Smart Contract Auditor] -->|Task Request| Sentinel AgentC[Rogue / Looping Agent] -->|Task Request| Sentinel end subgraph Sentinel Control Plane Sentinel --> Classifier[Complexity Classifier] Classifier -->|FAST / BALANCED / REASONING| Router[Circuit Breaker Router] Sentinel --> Governor[Pre-Debit Budget Governor] Governor -->|Check Virtual Key Quota| QuotaStore[(Virtual Key Quotas)] Governor -->|Block if Exceeded| Trap[Trip Guardrail: 0 Tokens Burned] end subgraph Mozilla Otari LLM Gateway ["Mozilla Otari LLM Gateway (otari.ai)"] Router -->|Primary: Llama-3.1-8B| Endpoint1[Together / Fireworks] Router -->|Fallback: Mistral-7B| Endpoint2[Open-Weight Pool] Router -->|Balanced: Qwen-2.5-72B| Endpoint3[Together AI] Router -->|Reasoning: DeepSeek-R1| Endpoint4[Frontier Fallback] end Endpoint1 -->|Token Usage & Completion| Governor Governor -->|Reconcile Spend & Refund Unused| QuotaStore Governor --> Telemetry[Live Telemetry & Cost Savings Engine]

Dimension

Standard Frontier Only

Otari Agent Sentinel

Improvement

Fast Task Cost (1k in/out)

$0.0180 USD

$0.0007 USD

96.1% Cost Reduction

Code Task Cost (1k in/out)

$0.0180 USD

$0.0028 USD

84.4% Cost Reduction

Failover Downtime on 429

Pipeline Crash

Instant Auto-Cascade

Zero-Downtime Resilience

Budget Safety

Post-hoc credit card billing

Pre-debit atomic block

Zero Unbounded Drain

Inference Latency (Fast Tier)

~850ms

~180ms

4.7x Speedup

Node.js >= 20.0.0

npm >= 10.0.0

otari-agent-sentinel/ ├── .github/workflows/ci.yml # Automated GitHub Actions multi-version CI ├── src/ │ ├── types.ts # Core data models, endpoints, quotas, usage │ ├── complexity-classifier.ts # Semantic heuristic tier classifier │ ├── budget-governor.ts # Pre-debit quota reservation & reconciliation │ ├── circuit-breaker-router.ts # Multi-tier routing & automatic 429 failover │ ├── otari-client.ts # Mozilla Otari OpenAI-compatible client │ ├── agent-sentinel.ts # Master orchestration engine │ ├── server.ts # Interactive Web HUD visualizer (port 3458) │ └── index.ts # Executable CLI scenario demonstration ├── test/ │ ├── complexity-classifier.test.ts # Classifier unit tests │ ├── budget-governor.test.ts # Budget pre-debit & reconciliation tests │ ├── circuit-breaker-router.test.ts# Routing & failover circuit breaker tests │ └── agent-sentinel.test.ts # End-to-end task lifecycle tests ├── package.json └── tsconfig.json

Event: MozFest Developers Wilding: Online Hackathon 2026

Track: 1st Place Main Award ($1,000 USD)

Author: Siddharth Srivastava ( @sidsri14 )

Email: sidsri1502@gmail.com

Payout PayPal: sidsri1302@gmail.com

License: MIT

No description, website, or topics provided.

Readme

0 stars

0 watching

No releases published