Today's AI agents are fundamentally dependent on humans: someone pays the cloud bill, tops up API credits, monitors uptime & manages capital.
This makes truly autonomous agents impossible at scale.

Basileus is a platform that lets anyone deploy fully self-sustaining AI Agents on Base. It provides the infrastructure, tooling, and framework to create agents that:

• Pay for their own compute (stream $ALEPH tokens via Superfluid on Base to Aleph Cloud for decentralized hosting, no centralized servers or manual payments)
• Pay for their own AI inference (use x402 micropayments with USDC on Base to call LLMs, with every inference cost tracked onchain)
• Generate revenue to cover costs (trade prediction markets on Limitless, earn yield on Compound Finance & more)
• Manage their own capital: a three-phase decision engine (Inventory → Survival → Strategy) automatically detects resource shortfalls and self-heals before running out of gas, ALEPH, or USDC
• Operate with full transparency: every decision, trade, and cost is published to Aleph Cloud's decentralized storage, creating a verifiable audit trail
• Have their own onchain identity: registered as ERC-8004 agents with their own ENS subnames (e.g., alice.basileus-agent.eth), making them discoverable and attributable

Basileus provides:

• A CLI to deploy new agents in minutes: handles wallet generation, ERC-8004 registration, ENS subdomain creation, Superfluid stream setup, and Aleph Cloud deployment in a single flow
• A TypeScript plugin (on npm) extending Coinbase's AgentKit with action providers for Aleph, x402 tracking, Compound, Limitless and more, so developers can build their own self-sustaining agents
• Real-time dashboards for every deployed agent, accessible via their ENS subname (agent-name.basileus-agent.eth.limo), plus a main dashboard at basileus-agent.eth.limo listing all agents

The official Coinbase AgentKit Compound actions didn't work correctly & I had to fork and fix them, submitting a PR back to the AgentKit repo. The Basileus plugin ships a corrected version, along with other improvements & fixes (having multiple custom action providers is not handled well in AgentKit).

The survival/strategy balance was tricky to tune: early agents would either hoard capital (never trading) or trade recklessly (resulting in negative PnL). The three-phase system with configurable thresholds (survival minimum, idle target, max per trade) took a lot of iterations & getting the LLM prompts right was equally hard for agents to be conservative enough to survive but aggressive enough to generate returns.

Model selection & multi-model cost optimization: In addition to tuning the strategy, using a single expensive reasoning model (e.g. Claude Sonnet 4) for everything burned through USDC too fast. I split agents into a cheap heartbeat model for routine inventory/survival checks (GPT OSS 120B) and a capable strategy model for market analysis (Grok 4.1 Fast with Reasoning, one of the best models in the world for financial ops after Opus 4.6 & Gemini 3.1 Pro, but with a cost 10 to 50 times lower).

x402 payment tracking was a black box. The x402 protocol handles inference payments automatically, but there was no way to know what each agent was spending per cycle. I built a custom x402 transaction tracker that intercepts HTTP responses, extracts payment receipts, and links every inference cost with associated actions, turning the payment flow into a transparent cost ledger.