Well, it's been a minute. I took an embarrassingly long hiatus from writing this newsletter. I did manage to write a book, but that’s no excuse. More on that next month. In the meantime, I'm back, and I have a lot to share.
In our March 2025 edition, I wrote about AI Agents - what they are, what they can do, and where they're headed. I painted a picture of a future where I'd have my own personal AI assistant, managing a team of specialty agents across work and life. I ended that piece with, "What a time to be alive."
Well, the future showed up faster than I expected. I’ve named her Juniper. And she's been on the job for about a month.
The Agents Are Here
Juniper is an AI agent based on an open source framework that lives on a virtual machine, communicates with me entirely through Slack, is connected to over a dozen of my business systems, and takes real action on my behalf. She reads emails, manages my calendar, sends contracts, reconciles my books, generates leads, audits cloud infrastructure, and as of today, responds to my voice through Siri on my phone. In four weeks.
I want to walk through what that actually looks like, because I think there’s an enormous gap between what people imagine agents can do and what they can actually do right now. And it's not in the direction you’d think.
What Juniper Has Done in 30 Days
Here's a sampling, and I'm leaving some top secret tasks out:
DocuSign Automation
If you've ever sent a contract through DocuSign manually, you know it's an annoying process. Log in, upload the document, add the recipients, place the fields and signature tabs, send it, then do it all again. Juniper set up API-based contract signing so that sending a two-party agreement - is now a single command. We've sent multiple live contracts this way. This was my first big win.
Automated Lead Generation
A client had an idea for a way to generate targeted solar leads based on a pretty novel strategy. The pipeline they needed was complex - multiple data sources, geocoding, property classification, enrichment with owner contact information, and daily automated reporting. It would have taken me a while to write that code myself. Juniper wrote over 1,500 lines of code in minutes, iterated through two major versions, and now the pipeline runs daily on autopilot, delivering verified leads to a Google Sheet and emailing the report to the client.
Monthly Financial Reconciliation
She connected to our QuickBooks Online account, built a reconciliation workflow that pulls revenue by client, project expenses, P&L line items, and payroll, then cross-references it all against my financial tracking spreadsheet. She completed our first real monthly reconciliation, mapping client names between systems, handling unpaid invoices that roll forward, and filling in detailed expense line items.
And the Rest
She connected with our HubSpot CRM, connected to Harvest for time tracking and invoicing across all clients, audited our Azure cloud infrastructure and found over $1,000 per month in savings, executed the cloud infrastructure changes, migrated our AI model routing to AWS for data privacy, and found a security vulnerability where the master password was sitting in a world-readable file (yikes). That last one was an interesting conversation.
It Doesn't Always Work
I'm not going to pretend this is all sunshine and rainbows. It's not. Agents are powerful but they are not perfect, and sometimes the failures are spectacular.
The most recurring issue is that Juniper will occasionally make a change to her own configuration, tell me something like "Quick update and a restart, and we're back in business," and then completely brick herself. No response. Gone. The first time this happened I panicked, because I had started to rely on her to fix things like this. Then I remembered I'm an engineer, so I SSH’ed into the server and figured out what went wrong. It's happened more than once.
The lead generation pipeline silently crashed on its first automated run. It processed 507 outages, found 148 leads, and then failed during the final write to the Google Sheet. The way the script was piped masked the error code, so there was no indication that anything had gone wrong. It just looked like nothing happened, yet it spent $150 on credits and tokens.
She's sent emails to the wrong recipients because a command line flag silently overwrote the first email address with the second. It took two days of the client not receiving reports before we caught it.
She'll burn through API credits on bad data. She'll confidently describe a fix that makes things worse. She requires supervision.
But here's the thing - so do people. Every junior employee I've ever managed has bricked something, sent an email to the wrong person, and confidently described a fix that didn't work. The difference is that Juniper works around the clock, never forgets what you told her (as long as she writes it down), and costs a fraction of a salary. The ROI, even with the failures, is absurd.
Captaining
So what does my workday actually look like now? It's fundamentally different than it was a month ago.
I run multiple agents, each dedicated to a specific domain - command center, marketing, finance, cloud infrastructure, sales, and more. Each lives in its own Slack channel. Same personality, different expertise.
My day has become just cycling through those channels. I open Slack, check in with the first agent, review what's been done since my last message, push the task forward with the next prompt, and move to the next channel. By the time I've cycled back to the first agent, she's completed the work and is ready for the next step. Any time not spent doing this is working on standing up a new agent for a new project or task and adding it to the rotation. Meetings - haven’t figured out a solution for. Yet.
I'm calling this Captaining. And I think it's the future of knowledge work.
The mental model is not "prompt engineering." It's not typing clever instructions into a chatbot. It's managing a team. Setting direction. Making judgment calls. Reviewing work product. Pushing things forward. The agents do the research, build the systems, handle the grunt work, and flag when something needs a human decision. I set the heading and manage the crew.
With this approach, I can now do the work of roughly 1.5-1.75 people. And honestly, that number feels like it's growing every day. I'm not working harder. I'm not working longer hours. I'm just checking in with my crew and keeping the oars moving in the same direction.
Four weeks ago I had an idea and an empty Slack workspace. Today I have a group of agents that are handling a lot of my typical daily tasks. It’s not perfect. It requires a captain. But it is here, the value is real, and it is transformative.
If your job involves reading, writing, researching, updating systems, managing data, sending emails, or performing any reasonably repetitive knowledge work, this is what your work will look like by the end of the year.
Bytes
- "I consider it malpractice to not use AI on my client work" - a very good lawyer who shall remain nameless.
- Perplexity launched Personal Computer, turning a Mac mini into a 24/7 AI agent. Sound familiar?
- BCG coined "AI brain fry" - workers managing too many AI tools saw 14% more mental effort and more errors. Productivity dipped after three tools.
- Morgan Stanley warns a non-linear jump in AI capabilities is coming by mid-2026. Most of the world isn't ready.
Last but not least, I started a podcast. Listen here.
