Prepared for Olam · Dolphin AI
The autonomous ops layer — already running.
Not a pitch for what I could build. A walk through the self-running system I operate for my own business right now — and how the same loops point at yours.
Paul Cowen · AI Orchestration — 1 June 2026
The distinction that matters
Sidekick vs. autonomous loop.
Workflow chain
A human triggers each step. A deterministic graph in n8n / Make / Zapier. Last year's "AI copilot" — a 20–30% productivity bump. Useful, but it stops the moment you stop.
Autonomous loop
The system observes, decides, acts, evaluates its own output, repairs itself, and keeps going on its own clock. The human is removed from the loop — it's already improved by the time you're back.
Your two videos draw exactly this line. It's the side I build on — and the bit most "AI automation" people skip.
What I ran this week — hands-off
A self-improving knowledge pipeline.
Point it at a course folder. It transcribes every video, extracts teachable frameworks + a domain ontology, embeds the synthesis, files it, and makes the cross-course graph richer for the next run. Each course makes the next forge smarter.
7
courses processed this week
190+
videos transcribed locally
35
frameworks + 7 ontologies forged
~6h
longest unattended run, zero supervision
The "while you sleep" economics
Multi-machine cost-routing.
Expensive intelligence sits only at orchestration and final judgement. Bulk labour routes to free local models across a private mesh — running 24/7 at zero credit cost.
| Job | Routed to | Why |
|---|
| Orchestration & judgement | Claude Opus (conductor) | The one brain worth paying for |
| Transcription (190+ videos) | faster-whisper, local M4 | Free, parallel, resumable |
| Framework + ontology forging | Codex / GLM-4.7 on "the Beast" | Bulk reasoning, free, 24/7 |
| Failover when a box stalls | 48GB node over Tailscale | Live multi-machine recovery |
Built last night · the part agencies actually need
Least-privilege agent access.
Autonomous agents reading a live database is where most setups quietly leak. Mine read a curated slice through scoped keys — denied at the Postgres grant layer on everything else, including where row-security is off.
| Agent | Course library | Full text + search | Payments · clients · memory | Writes |
|---|
| Hermes (tightest) | ✓ read | ✕ denied | ✕ denied | ✕ denied |
| OpenClaw (hardened) | ✓ read | ✓ read | ✕ denied | ✕ denied |
Never the master key. Curated views, standalone roles, revocable in one line. Exactly the governance an agency's "policy layer" demands.
The architecture your video names — in my stack
Five layers, live.
Sensor
Emails, transcripts, Slack/WhatsApp, telemetry ingested as data.
Policy
What runs autonomously vs. what needs sign-off — and least-privilege access, logged.
Tool
Deterministic skills & APIs — Supabase, model CLIs, 400+ skills.
Quality gate
Eval + safety; irreversible actions (send/deploy/delete) hard-gated to approval.
Learning
Friction fed back to the top; new skills & memories persisted. The loop closes.
For Dolphin AI
How we'd start.
- Map your Sensor layer first — what's already captured as data (Slack, email, meetings, telemetry) sets how fast we stand up the "company brain".
- Point the same loops at your repo — agent → identifies gap → writes tool → opens PR → deploys, behind the gates you choose.
- Pin the autonomy dial together — how much ships overnight vs. waits for a human. That decision scopes the whole build.
- Land one ACE loop end-to-end — prove it on a real, high-value workflow before widening. Quick win, then scale.
The honest gap
My production autonomy runs in my harness today — not yet inside a client repo with merge-and-deploy loops like your first video. That's an integration job, not a capability gap: same loop, your codebase, your deploy gates. I'd rather flag it now than have you find it.
If that's the shape of what you're building — let's wire it to your stack.