I get asked what I actually do with AI agents quite a bit. I work with them full time while helping manage several companies, coaching clients and keeping up with personal projects. Most of it is the fairly ordinary work around those things. Here’s what I have mine do on a regular basis.

• Billing and email. Investigate problems, check related records and conversations, and bring me the facts with suggested next steps.
• Business reports. Gather financial and performance data, flag concerns and prepare weekly reports we can discuss in Slack.
• Marketing. Research product updates, trends and interesting stories, then draft newsletters and social posts for review.
• Support. Investigate tickets, reply when the issue is low risk and the answer is clear, and escalate other cases to a human.
• Development. Work through assigned tasks using a defined process with evaluations and audited steps.
• Meeting prep. Research people I haven’t met, look back over previous conversations and prepare context before a call.
• Calls and bookings. Call businesses to arrange appointments, then keep the result and a transcript for review.
• Research and notes. Turn meeting notes and saved videos into useful reference material, and keep people and project notes current.
• Workspace housekeeping. Check that files follow our organization rules and report anything out of place.
• Personal errands. Work through things like our grocery history and this week’s sales to prepare an order for me to review.

Some of this starts with a task I give the agent, while other work runs on a schedule. The useful part is how the steps connect. An agent can find an issue, investigate it and either deal with it within its instructions or bring back enough context for a person to decide what happens next.

Keeping up with the companies

One of my recurring jobs checks the coworking business’s billing each day. It looks for overdue invoices, failed payments and membership changes. If a member has both an overdue invoice and a subscription marked past due, it groups those together so we can look at the actual issue.

Before coming back to me, it checks the payment history, reads related email threads and gathers the context around what’s happened. That investigation is part of the job. I get the key facts, relevant history and suggested next steps in Slack, where I can review them and give it direction. We can prepare an email or arrange a follow-up from the same conversation.

This is where the little steps become pretty damn useful. Checking the billing system saves one bit of work. Connecting that to the email history and checking what we’ve already discussed means I can start thinking about how to handle the situation. I don’t need to ask it to go back and investigate the thing it just told me about.

Email works much the same way. I have jobs checking business inboxes, reading threads and looking up relevant context before bringing me something that needs attention. Newsletters and routine updates can stay quiet. I don’t need a second inbox containing a summary of every message in the first one.

I also need to keep an eye on how the companies are doing overall. For each one, my agent gathers information from Xero and the platform’s metrics tools to track finances and key performance indicators, the measures we use to judge how the business is doing. It brings that information into a weekly report in Slack, with the trajectory and any concerns for us to work through.

From there, I can ask about a change, check the assumptions behind it and dig into the relevant details. Those reports help me keep a view across the companies and work with the teams on what each quarter’s direction and focus should be.

For example, I’ve used agents to compare the agency’s scheduled project work with its billing history. That helps me see when booked work runs out, who has room for another project and what that could mean for revenue. The schedule and the accounts each tell part of the story. Work that hasn’t been entered yet, or an old placeholder that nobody removed, can change the picture quite a bit, so those assumptions still need checking.

Marketing, support and development

For some of the projects I help run, we have dedicated agents doing marketing and support work. The marketing agents research recent product updates, relevant trends and interesting stories or links, then use that material to draft newsletters and social posts for us to review. Refining the voice has been the hardest part. Getting the information together is useful, but the writing still needs to sound like the people and the product it represents. That’s something we’ve had to keep working on.

Our support agents look for new tickets, work out what the issue is and gather the key information. If the issue is low risk and there’s a clear answer, they can respond directly. Otherwise, they let the person know we’re looking into it and pass the issue to a human with the context they’ve collected. The person taking over has something to work from, and the customer knows their question has been picked up.

I also feed tasks to a development agent that works through a pipeline, a defined sequence of steps with evaluations to check the work along the way. Those steps are audited, so there’s a process I can inspect as it works on a project.

Knowing who I’m meeting

My daily prep brings together today’s calendar, recent notes and material relevant to the meetings ahead. If I’m meeting someone new, the agent does some background research and creates a note for them in our shared Obsidian knowledgebase. I can read about who they are and what they’re working on before we get on the call.

For someone I’ve met before, there’s already a history to work with. The agent can find our earlier conversations, the project we’re discussing and any decisions that matter for the next meeting. That helps me get back into the subject without having to reconstruct it from calendar entries and old messages.

Afterwards, another job takes the notes from my meeting tools, pulls out decisions and next steps, and files them with the right company or project. The person’s note gets updated with useful details and dates as we continue working together. Those dates matter because something we discussed six months ago may no longer describe what they’re doing now.

The next briefing can use that history, along with research we’ve saved between calls. Each conversation gives the next one a little more to work with, provided we take care of the notes.

Calls, bookings and the groceries

I use agents to call businesses and arrange appointments too. I give the agent the purpose of the call and the details it needs, including limits like which times work for me. For a restaurant booking, that means the party size, preferred times and what to ask if those aren’t available.

It makes the call and brings back the result, with a saved transcript so we can review the conversation afterwards if needed. Sometimes it gets voicemail or can’t get one of the requested times. That comes back as something still to deal with, rather than a booking I can assume is finished.

The same arrangement is useful for personal errands. I’ve had my agent look through our grocery order history, work out what we usually buy, check the sales and prepare the week’s order for me to review. Fairly unglamorous, but I’ll happily take one less thing to organize.

Giving the work somewhere to live

I built Enso to connect the agents running on my computer to Slack and handle the background jobs and follow-ups. Each Slack channel is connected to a workspace, a directory with instructions, reference material and somewhere to put the work. I have separate ones for the companies, development and general life stuff.

The coworking workspace knows about memberships, billing and how we handle enquiries. When I ask about an invoice there, I shouldn’t have to explain which company we’re talking about. The relevant instructions and skills live close to that work, and I can open the channel to see what’s happening with it.

I also have an agent check the workspace against our organization rules and report anything out of place. Those rules explain where work belongs and which files are authoritative. I love a source of truth. It gives both of us somewhere to look, and something to check against when files start turning up in odd places. I wrote more about that setup in Your agent is a slob.

A workspace organizes the context; it doesn’t restrict the agent’s access by itself. Shared preferences still apply, and the computer needs to be running for the agent to do its work. Through Slack, I can read an update, ask a question and approve the next step from my phone while the files and tools stay on the computer.

Work that continues between conversations

A job is a recurring responsibility with instructions and a schedule. Once I’ve set it up, I don’t need to send the same request every morning. Billing checks, weekly reports, meeting preparation and housekeeping can each run when they’re useful.

They don’t all need to interrupt me. The daily prep can be waiting in my notes, while a billing problem belongs in Slack. If a check finds nothing new, I’d prefer it to say nothing. Automating a daily message telling me there’s no daily message would be a fairly stupid use of everyone’s time.

Alongside those jobs, I can ask for a temporary follow-up with a stopping point. Enso calls that a heartbeat. I’ve used it to watch for a reply about an invoice, bring the answer back to the Slack thread where we’d discussed it, and then stop watching. Waiting for one reply doesn’t need to become another permanent job.

That follow-up is recorded with its timing and progress, so it doesn’t depend on me keeping the chat open. An agent saying it’ll keep an eye on something isn’t much use unless something is actually going to wake it up later.

Teaching it something we can use again

I wrote earlier this year about using Obsidian as a second brain and having an agent help organize and use my notes. Obsidian stores them as plain text files, so both the agent and I can read and update the same material. That shared knowledgebase holds research, meeting history, people and project decisions.

I also have a job watching a YouTube playlist I use for things I want researched. It’s set up to turn new additions into notes using the transcript and supporting sources. Research saved there can become context for a later project or conversation, alongside the material we gather through our own work.

The useful details need a clear home. A decision buried in a long chat is easy to lose when another agent starts working on the same thing. A note with the decision, its source and a link to the project gives it somewhere specific to look.

Skills carry the procedures forward. A skill is a set of instructions for doing a particular kind of work, sometimes with scripts or reference files alongside it. I have them for email, billing, research and working with various tools. When an agent struggles, we can work through the problem together and save the procedure that worked.

That’s the part of compound engineering I find useful here. Doing the work also gives us a chance to improve how the next piece of work gets done. Once we’ve figured out a process, we don’t need to start from scratch every time it comes up.

I still want a hand in what becomes a skill. Letting the agent keep generating procedures on its own can leave you with a pile of instructions that overlap or disagree. We need to check what already exists, decide where the lesson belongs and refine what’s there. Sometimes that means removing an instruction that is causing trouble.

I’ve seen this in coaching too. One client’s agent was picking up broad skills instead of the intended project-specific ones. We reorganized the work and tightened the instructions so it was clearer which process belonged to that project. Adding more instructions wouldn’t have addressed what was already being loaded.

The same care applies to the rest of the setup. Integrations break, accounts change and agents still put things in the wrong place. I’ve had to fix a meeting-note job that was looking in the wrong account and finding nothing. Saving a process gives us something to reuse and refine, but I still need to check that it’s doing the right thing.

If you’re looking for a useful first task, I’d pick something you already do regularly. Give the agent the relevant history, have it prepare a result you can check and work through what it gets wrong. Once that works, decide whether it should run on a schedule and where the result should go. Keep the procedure focused enough that you can understand and maintain it. You’ll get more out of one job you can rely on than a folder full of skills you haven’t read.