Research
Research: AI-Operating Knowledge
A modular, self-updating research surface.
Seven topics on running AI inside a real business: memory, knowledge flywheels, orchestration, skills, governance, costs, and the open-vs-closed-model question.
Each module is a standalone page. Each carries the date it was last reviewed and a short note on what that edition covers, so you can see how current it is at a glance.
Subscribe to the modules relevant to your work. Skip the ones that aren’t.
The whitepaper
Countersign: Governing Agentic AI by Construction
A whitepaper for the people who will adopt AI agents and the people who will regulate them.
Working draft, June 2026. Revised July 2026.
For years, governing AI meant governing advice: a person made the final decision, so the rules shaped that person's judgement. AI agents change this. An agent acts. It calls tools, writes to systems, moves money, and it does so at machine speed with authority someone delegated to it. A written policy does not stop the action it forbids, and an agent asked to police itself is judging its own case. The paper sets out a different arrangement: the agent does the work, a named person stays accountable for it, and that division is enforced by a small piece of code outside the agent that the model cannot rewrite, persuade, or skip. Every promise the framework makes is stated, tied to the control that enforces it, and backed by a test that fails if the control is removed. The paper grades its own promises, says which are finished and which are still to come, and is clear that its testing so far is its own.
- The executive briefOne page. The problem, the arrangement, and where the work stands.
- The Singapore field guideWhat IMDA, MAS and PDPC expect, read as a practical guide for someone deploying agents here.
- Ten questions for your agent deploymentA self-check to run before you switch an agent on. A no is worth knowing early.
- Cross the gateA short presentation of the core idea: the decision is made outside the agent.
The paper is still in revision. Leave an address and I will send it when it is out.
The modules
All seven modules are live. Each is a standalone, research-backed page with a “what this edition covers” summary at the top, so you can scan for what’s new without re-reading what hasn’t moved. They are independent. Read them in any order.
- AI Memory. The working/persistent boundary, RAG’s limits as memory, and what memory costs at scale.
- The Knowledge Flywheel. How knowledge accumulates in operated systems: the documentation → agent → execution → learning loop, and why flywheels stall.
- Orchestration. Multi-agent coordination: orchestrator, swarm, and hierarchical topologies, and what actually fails in production.
- Skills + Tools. How agents are equipped with capabilities: tool use, the Model Context Protocol, skills, and tool-use governance.
- Governance. Audit trails, escalation thresholds, human-in-the-loop discipline, and the 2026 regulatory picture (EU AI Act, Singapore’s agentic framework).
- Costs. Token economics, the inference cost curve, and operational cost discipline for a system running thousands of calls a month.
- Open vs Closed Models. When to use closed (Claude / GPT / Gemini) vs open-weight (Llama / Mistral / DeepSeek), and the self-hosting break-even.
How the Research works
Modular. Each module is its own page. Read one; skip the rest. No 15,000-word scroll.
Dated. Every module carries the date it was last reviewed. We refresh modules as the research moves, model releases, new benchmarks, regulatory changes, rather than on a fixed treadmill.
Sourced. Every module is research-backed, and the load-bearing figures carry citations to their primary sources.
Yours to choose. Read only the modules relevant to your work. They’re independent, in any order. Follow new editions by RSS, or get the monthly Notes digest.
Research is how we think out loud about running AI in a real business. The category-definition essay AI-operated vs AI-generated is the companion piece. It names the position the Hub’s research sits inside.
Related
Get in touch
Stuck on something?
Tell me what it is. I read every message myself and reply with what I would do about it.