The AI Operating System
Five layers deep. 54 apps. One governed fleet — receipted, verified, and cost-attributed. Built by Knox. Running 24/7.
Five Operating Layers
The Agent Layer
The orchestrator receives directives, routes tasks to Directors, and spawns Claude Code worker sessions via skills. It never writes files — it prompt-engineers downstream. The Principal-Broker manages model-tier routing: Haiku for observation, Sonnet for planning, Opus only when a directive is generated. The Advisory Council — a seven-agent E-Board — operates within this layer. CFO, CTO, CMO, Risk, Strategy, CRO, and CDO each hold autonomous authority within defined mandates. The Chief of Staff triages escalations and produces Knox's daily digest. Knox reads one brief. Approves three items. Issues zero to one directives. The company runs.
The Intelligence Layer
Akashic Records is the memory system. 41 categories. Thousands of entries spanning trading observations, engineering lessons, content strategy, and project history. 92% codebase coverage across 59 repositories. Agents read Akashic before every session — mechanically enforced, not instructional. Each agent also maintains a self-updating expertise file that compounds its learnings between sessions. The fleet gets smarter without Knox. The InDecision Framework scores markets with six factors and 82.5% directional accuracy. Tesseract Intelligence synthesizes competitive signals into structured briefings. Not summaries. Synthesis.
The Build Layer
Claude Code builds new apps, fixes bugs, and ships features autonomously. 54 apps and counting — all built through this pipeline. Knox directs via Discord. Claude Code executes. Mission Control tracks the output. The skill library governs what workers can do: 66 skills covering every build pattern, each with a 90-day freshness window before drift governance flags them for review.
The Monitoring Layer
Invictus Sentinel watches every service, tunnel, and cron job. The API credit preflight checks balance and burn rate every six hours — on first deployment it returned a live WARN condition before a single session was killed. Token ledger attributes every API call to a specific agent and model tier. Skill health tracking logs every invocation across all 66 skills. Nothing runs blind. Nothing fails silently. The watchers are never the watched — monitoring runs via launchd, independent of the agents it monitors.
The Governance Layer
Most AI operating systems trust their agents. This one verifies them. The Harness is the enforcement layer running underneath all 54 apps. Six mechanical gates that turn a capable fleet into a provably reliable one. Not a framework. Not a convention. Infrastructure with receipts.
SEE THE HARNESS →
Akashic Records
The knowledge base that governs everything the fleet knows.
41 categories. 1,496 knowledge-base entries. 852 journal entries. 668 lessons. 221 trading observations. 63 content strategy documents. 92% codebase coverage across 59 indexed repositories.
Akashic is not a database agents query on request. It is the mandatory context layer agents must read before work begins. The Harness enforces this with a cryptographically-keyed read receipt at every dispatch. Akashic read + receipt written: work begins. No receipt: blocked.
The per-agent expertise layer sits on top of Akashic. Each agent — Advisory Council, Foresight, Apollo, Shiva, Sentinel — maintains its own expertise file that it self-updates mid-session. By session ten, Advisory Council's expertise file contains observations about the fleet that Knox never wrote. The system learns without being taught.
Akashic Records powers ask.jeremyknox.ai — the AI Knox that answers questions as Knox would, drawing from 14 years of compounded knowledge.
What Tesseract Has Built
Built from 7 years of market analysis. Six factors, custom-weighted, producing directional conviction scores on crypto assets — now powering autonomous Polymarket trading.
Always-on monitoring stack. Detects failures across services, tunnels, and pipelines. Fires Discord alerts with diagnostics. Built after a 1-hour outage to ensure it never happens again.
The command center for the entire Tesseract Labs ecosystem. Live service health, cron status, cost tracking, and quick-action shortcuts — accessible from any device via cloudflared tunnel.
Seven specialized agents — CFO, CTO, CMO, Risk, Strategy, CRO, CDO — each operating within a defined authority matrix. Auto-execute within mandate. Escalate only what requires Knox. Powered by per-agent expertise files that compound session over session.
The governance layer for the entire fleet. Built April 13, 2026 in a single swarm session: credit preflight, read receipt enforcement, independent completion verification, cost attribution, skill drift governance, and peer agent bus. All six items shipped. All 77 tests passing.
Under the Hood
The Philosophy
Gather → Synthesize → Deliver
Every pipeline decomposes into three stages. No monoliths. Each stage has its own timeout, blast radius, and failure mode — like Netflix microservices applied to AI automation.
Script-First, AI-Last
If a task can be a Python script, it stays a Python script. AI is reserved for synthesis, analysis, and judgment calls — not data fetching. Keeps costs low and reliability high.
Fire-and-Forget Delivery
AI writes content to files; scripts post to Discord. AI generates analysis; crons distribute to Signal. Decoupled delivery means one component failure never kills the whole pipeline.
Alerting Independent of the System
Sentinel and the tunnel watchdog run via launchd, independent of Soul. If the agent goes down, monitoring stays up. Knox still gets alerted. The watchers are never the watched.
Governance Before Trust
Agents don't get credit for completing work — the harness verifies it. Agents don't get trusted to read context — the harness receipts it. Agents don't get assumed to stay within budget — the harness attributes every call. Trust is what you have before a system matures. Governance is what you build after.
Expertise Compounds
Every agent session ends with a write. What the agent learned, what failed, what it would do differently — written to a persistent expertise file injected at the next session start. Akashic Records is the fleet's shared memory. Expertise files are each agent's private memory. The fleet gets smarter every night without Knox in the loop.
The Harness
Most AI operating systems trust their agents. This one verifies them.
The Harness is the enforcement layer running underneath all 54 apps. Not a framework. Not a pattern library. A set of mechanical gates — each with tests, each with receipts — that turn a capable agent fleet into a provably reliable one.
Built in a single swarm session on April 13, 2026. 77 tests passing before midnight.
CREDIT PREFLIGHT
Before any Opus-class session initiates, the harness queries API credit balance and projects daily burn rate. On first deployment, it returned a live WARN condition: $5.43 remaining at $10.91/day burn. Caught before it killed a session. Silent failures become loud alerts.
AKASHIC READ RECEIPT
Every agent must read its knowledge base before work begins — not as an instruction, as a gate. A read receipt keyed to the session ID is required at dispatch. No receipt: blocked, logged, alerted. One retry. Then escalated to Knox. The constraint is mechanical, not instructional.
INDEPENDENT COMPLETION VERIFICATION
When an agent marks a task done, a separate Haiku-class verifier checks the output against acceptance criteria before it archives. Partial completion triggers one retry with failing criteria surfaced. Second failure escalates to Knox. Agents don't grade their own work.
PER-AGENT COST ATTRIBUTION
Every API call is attributed to a specific agent and model tier. Haiku for observation. Sonnet for planning. Opus for directives only. Daily attribution reports show top agents by spend, tier breakdown, and 7-day burn rate. The cost model is enforced, not assumed.
SKILL DRIFT GOVERNANCE
66 skills entered a 90-day freshness window on April 13, 2026. Any skill without a successful invocation in 90 days is flagged [REVIEW]. Not deleted — surfaced. Autoresearch processes flagged skills first. The library compounds. It doesn't rot silently.
PEER AGENT BUS
The E-Board routes peer-to-peer without Knox as broker. CMO requests data from CRO directly. Strategy queries CDO for calibration state. The harness logs every exchange to the agent bus event timeline. Knox only sees ESCALATE message types. The company deliberates without him.
This isn't infrastructure we built to sell. It's infrastructure we built because the fleet demanded it. Now it's both.
See All 54 Apps
The full arsenal — every app, every pipeline, every system. Eight categories of autonomous power. Governed by one harness.