Engine

Origin

Most production AI systems are single-agent: one model, one prompt, one task. That works until you need a system that runs unattended for weeks, holds opinions, and learns from its own failures. Engine was the bet that an autonomous civilization of agents — with hierarchy, lifecycle, prediction, and self-modification — would beat any single-shot orchestration.

Problem

How do you get hundreds of agents to cooperate without a central planner? How do you let them propose changes to their own architecture without losing the ability to roll back? How do you keep them honest when they grade their own predictions? Engine's answer is structure, not eloquence: hierarchy, lifecycle states, calibration loops, and explicit self-modification proposals.

Approach

Four empires — covering AI, marketing, quantitative finance, and crypto — each with its own lieutenants and operating cadence. Lieutenants spawn heroes; each hero walks a deterministic lifecycle and dispatches small specialist sub-agents (scout, analyst, executor, monitor) for the work it cannot do alone. A central panel sits above the empires for cross-domain intelligence, judgment, and the morning brief. A five-tier memory system makes institutional memory survive any single hero's death. A tiered model router routes calls by stakes — cheap models for the routine, capable models for the consequential — measuring cost in cents per cycle, not per query.

Methodology

A relational substrate, not a vector blob. Every fact, decision, prediction, and correction is a structured row with explicit relationships, queryable in real time. Compounding loops run on top — source reliability is updated against outcome, predictions are graded on a fixed cadence and miscalibration propagates into the directive set, and a self-improvement pipeline produces and audits proposals to the system's own architecture. The emperor still approves; the system proposes.

Selected milestones

• Four empires operational across distinct domains
• Tiered scheduler with circuit breaker on repeated failure
• Brier-score calibration on a fixed predict→verify→update cadence
• Paper-trading lane validates strategies before any real capital moves
• Self-improvement pipeline driving the proposal lifecycle, with human approval at the boundary

Open questions

• How calibration drift propagates across distinct empires (markets vs. marketing vs. AI)
• What the right level of autonomy is for self-modification — proposal, draft, applied
• How to keep cross-empire intelligence federated without losing per-empire signal

Ask me about

• Why hierarchy beats flat orchestration for unattended runs
• How prediction calibration is wired into the scheduler
• What the self-improvement pipeline produces and what it deliberately cannot
• How the tiered model router decides who gets which call

Stack