A good manager of humans is a good manager of agents
Your Essential Report from the Frontlines of Applied AI
You are not just a user of AI. You can be a manager of AI — by setting up numerous agents that can handle parts of your job. Like your personal team of competent, tireless juniors.
So this week is about doing that job properly: brief it, limit how many you run, and check what comes back. A prompt below for the first, and a city IT desk that got all three right.
Plus the week’s news — officials and factory staff building their own AI tools, NHS doctors on how medical AI should be approved, and 26 Croatian public bodies running AI with no rules for it.
A good manager of humans is a good manager of agents.
Used only as a chatbot, AI is an upgraded Google search: you ask, it answers, you decide what happens next. To level up: set up agents to take over parts of your workflow. An example: your weekly follow-up with your suppliers — find every purchase order past its promised date, draft a chaser to the supplier, put the list in front of you. Work you could have hired a junior to do.
So set it up as you would that junior. Outline the steps. Set the limit of what it can decide alone. State the objective. Describe what a good result looks like. Specify how you want it to report to you.
Then let it work, and wait. It reverts when it accomplishes the goal you gave it, and then you check its work. Exactly how you would run a person — only now you can run 50 at once.
But just because you can, does not mean you should. When your mundane tasks are done by agents, your job evolves into managing the agents and checking their work.
As tempting as it may be to set up dozens of agents the first time, the number you can actually manage is the number whose work you have time to read and correct.
Start setting up your agents, get comfortable, and soon you can set up manager agents to check your worker agents. That next level up will put you on the path to having 50 agents doing work for you.
- Public sector
Korea’s procurement agency put seven AI services into live work, built by its own officials
A task force of 26 staff, running since February, used ‘vibe coding’ so officials who cannot code could describe what they wanted in plain language. The seven now in use cover procurement-law consultation, drafting press releases and official documents, parliamentary business and workflow automation. Commissioner Baek Seung-bo demonstrated them on 11 September.
Public Procurement Service of Korea, 11 September 2026
- Manufacturing
LG Display has staff with no coding background building their own tools, and has automated 190 of their jobs
Under AX SQUAD, employees pick a repetitive task in their own work and build the fix with AI tools, coached by specialists. Around 450 projects so far, roughly 190 of them now running as real automation, including defect classification on the factory floor. It is moving out of manufacturing into procurement, customer support, finance and sales. The counts are the company’s.
Seoul Economic Daily, 13 September 2026
- Healthcare
NHS doctors want medical AI approved in stages, then monitored for as long as it is in use
The National Commission into the Regulation of AI in Healthcare, led by NHS consultants Alastair Denniston and Henrietta Hughes, wants new models deployed under close supervision and tight guardrails before fuller authorisation, and continuous real-world monitoring afterwards rather than a single sign-off. It heard from more than 12,000 people over a year, and asks for meaningful human oversight and clear accountability when something goes wrong.
MHRA, 10 September 2026
- Governance
Croatian police, courts and ministries are running AI; 26 of them have no rules for it
Of 78 leading public bodies, 50 use or test AI, and 26 of those said they have no internal rules governing it — the first systematic map of AI across the country’s public sector, built by the watchdog Gong through freedom-of-information requests. The interior ministry runs facial recognition and biometric identification. The government itself, the Constitutional Court and the State Audit Office use none.
Gong, reported by Novi list, 10 September 2026
An SOP is the brief. Two of its five sections are the ones an agent needs most: the numbered steps, and the decision points and escalation triggers — where it stops and comes back to you. Write it as though a new colleague were going to run the process, because the test in the prompt is exactly that: clear enough to follow without further instruction. Then hand the same document to an agent. It is written for a banking process and works on any process you repeat.
Persona: You are a banking operations professional documenting a standard operating procedure. Request: Draft an SOP for [process name] covering: (1) process overview and objective, (2) scope and applicability, (3) step-by-step procedure (numbered), (4) decision points and escalation triggers, (5) documentation requirements. Intent: This SOP will be used by team members performing this process. It must be clear enough for a competent employee to follow without additional instruction. Structure: Five sections. Step-by-step section: numbered list, each step as a clear action. Maximum two pages. Process details (you provide — AI will structure): - Process name: [Name] - Objective: [What it achieves] - Applicable to: [Which roles / situations] - Process steps (your description): [Walk through the process as you know it] - Decision points: [Where judgment calls are required and what drives them] - Documentation required: [What records must be kept]
Prompt 6.5 from AI+ Bankers, word for word.
Give the agents the high-volume, documentation-heavy work, and keep the complex half with the specialists.
Raleigh serves about half a million people with a small IT team. It put a ServiceNow agent, Ral-E, on the internal service desk to document tasks and handle routine requests, and left the infrastructure decisions with the specialists.
95% of help tickets are now auto-summarised and approved, and tickets clear 66% faster, with the AI drawing on knowledge-base articles and closed-ticket history — reported by InformationWeek. Its chief information officer, Mark Wittenburg, says it saved ‘the equivalent of a full month of time’, in a ServiceNow press release, so that one is the vendor’s.
Note the word approved. The agent drafts the summary and a person signs it off — which is the difference between a queue that moves and a queue nobody is reading.
