I'd hired a dozen AI agents without a workforce plan

The AI Generalist course I'm doing with Outskill demoed a new AI agent and I wanted it straight away. They showed several during the course and every one looked useful, which is the problem with a good demo. I'd been asking Claude Code about a few other agent platforms I'm desperate to try, so before I committed we went through what I already had: what does what and why. It came to around a dozen.

By an agent I mean something that does a job on its own trigger, with access to my tools, without me prompting each step. A chatbot that only replies when I message it doesn't count. Each of the dozen was built for a reason, one at a time, over a few months, and I'd never looked at them as a whole. I came to this assessment as an HR person, and I think a lot of AI adoption problems come down to workforce planning. So I did what I'd do with a team I'd inherited: I audited it.

Agents arrive like unsolicited CVs

A lot of agents seem to get hired the way unsolicited CVs get interviews, and I've watched the original version happen. In a large professional services firm I worked in, it was common at one point for a recruitment agency to send a CV straight to a hiring manager. The manager would interview the candidate and want to make an offer, without recruitment being involved. Historically, partners there had a lot of control over hiring in their own groups, and nobody had asked whether the headcount was approved, whether anyone else should get the chance to apply, or what the candidate was being measured against. After a couple of those cases we set up a much more rigorous process, with proper approvals and direct sourcing, and as the firm became more centralised the old way stopped.

That gap is also where bias gets in. If the role isn't defined before the hiring starts and nobody else gets the chance to apply, the decision comes down to who the manager happened to meet and like the look of, and that's unfair to everyone who never saw the vacancy. Defined criteria and an open process fix that for people, and they do the same job for agents, because they stop you choosing on enthusiasm.

A course build and demo works the same way. It arrives polished and unsolicited, and because you build the agents in real time as you follow along, it makes a very good case for itself.

Some of my agents were hired against a need. My PA bot exists because I wanted one place to capture to-dos and diary events, by text or voice. Others were hired because the CV looked good, and a couple came out of a workshop build. Several of those have turned out to be useful, Lifebot most of all, the one that sends me a morning brief, an evening summary and a weekly review, but I hadn't put them through the questions I'd ask about a new hire: is there a vacancy, who is it for, and how will we know it's working?

The audit found overlap and roles that never started

Three assistants were doing what was really one job. A to-do bot, a personal bot and a daily AI news briefing were each sending me a morning message, and overlapping roles are the first thing a workforce review finds. I lined two up at 09:00 and 09:05 and told the second to skip anything the first had already sent. The news briefing still goes at 09:30, but only when there's one to send.

Some roles never started. Six jobs I set Lifebot up to do, such as pulling in my notes and calendar, have never run. Of my two automations, one ran once and I paused it, and the other has only been tested on a recording of a single meeting, though I still think it's worth keeping. In recruitment terms, that's an offer accepted by someone who never turned up, and I'd rather count those honestly than leave them on the headcount.

The twelve are grouped by job, and anything that isn't simply running is labelled.

Every stage of managing people has an equivalent for agents, and this is the map I followed.

Every agent needs a job description and a reason to exist

The test I use now is one line: which existing tool does this replace, or which job does it take over? If I can't write that line, there's no vacancy, and the agent goes on a throwaway account with a time limit, which is as close as I can get to a fixed-term contract.

Anything that stays has to pass four tests in probation. It works on real data, it has a real trigger, it beats simply asking Claude, and someone actually uses it. Fail two of the four and it doesn't stay.

The PA bot's job description fits in five lines. The owner is me, the job is capturing to-dos and diary events and sending the weekday list, the trigger is my messages and the 09:00 list, the review date is 30 November, and it goes if it fails two of the four tests.

Access should start small, and mine hasn't yet

You wouldn't give a new starter every system on day one. I narrowed Tutorbot's access on 31 October to a short list of tools, where before it had the whole standard toolkit. Lifebot still has the whole toolkit, including the ability to run commands on my machine, and the PA bot's calendar permission allows it to delete events, which isn't in its job description. Both are on my list, so the audit isn't finished.

Performance reviews need evidence, and some of it is awkward

A review has to look at what an agent has produced, not just whether it's switched on, because Lifebot's two faults only showed up that way.

The PA bot has the clearest record. It has sent the weekday list on every working day since it went live on 27 October, and between 24 October and 2 November it captured 14 to-dos, seven of them done. Lifebot is the one I find most useful for keeping track of things, but its Sunday review stopped when a sign-in expired, and this week its morning message arrived about two hours late because it runs on my laptop.

Tutorbot, which quizzes me on AI, HR and recruitment from my own notes, is still on probation. I narrowed its job on 31 October, its first scheduled quiz ran on 2 November, and it hasn't produced a score yet.

The applications sync has made the biggest difference, because it ties the job search into one system. The job scan finds vacancies in the background, they land in Airtable, and my phone tells me what's new. I look at the ones that seem a good fit and apply, the sync keeps my tracker in line with my applications dashboard, and if a role reaches interview stage, Interview Coach builds a prep pack. The sync has succeeded on all six of its runs.

Consider redeployment before redundancy, and run the exit properly

One role ended and one is coming back. Cartero was a daily inbox-triage task that last ran on 24 September and, according to my notes, couldn't reach the inbox it was set up for. I'm bringing it back and pointing it at my personal inbox, which has far more activity. Wikibot was a bot for an investing-wiki exercise that I never used, and that one has gone.

Before making a role redundant I'd ask what depends on it, and the same question applied to Wikibot. Tutorbot had a rule that created a task card and assigned it to Wikibot whenever the wiki didn't cover a topic, so removing Wikibot would have left Tutorbot sending work to a role that no longer existed. I took the rule out of Tutorbot's instructions as part of the exit.

The coloured arrow marks the hidden dependency.

I backed Wikibot up before deleting its profile, in case I was wrong. I haven't yet tested that I can restore it, so for now it's a backup in name only. I disabled Cartero and didn't delete it, so its instructions were still there when the role came back, and it goes through the same four tests as anything new.

Questions I ask before approving the next hire

Before any new agent gets approved, I want an answer to each of these:

• What need does it meet, who raised it, and who owns it?
• What does passing probation look like, and when is the review?
• What can it access on day one, and what does it cost?
• Who is affected if it goes wrong, and were they asked?
• What happens if it doesn't work out?

I've frozen new agent platforms for 30 days, which is a hiring freeze by another name, and the work until then is to finish the reviews I've started. An agent without an owner is a change without a sponsor, and in a team the same questions apply to every tool someone brings in, whether they found it themselves or a vendor switched it on. The Grok bot will have to wait until the end of November.

How many of the AI tools your team uses could you point to an owner and a job description for?