Post
Those Agents, So Hot Right Now
Building agents as business partners unlocked time for me to focus on what I am good at.
June 1, 2026
**
Why I Built This
I run this alone. One person, fractional data and AI consulting, cannabis MSOs and mid-market companies. I have had teams to support, engage and challenge me for YEARS. Now I don't, but I still need help
What I do have is a pile of recurring operational work that doesn't scale with my hours. Development projects, exploring and building new capabilities (like this one.) Meeting recaps. Research prep before every prospect call. Invoicing, project management. Maintaining the cannabis domain knowledge I pick up and then lose if I don't write it down somewhere.
This stuff matters. It's the difference between showing up prepared and showing up reactive. It also eats hours I'd rather spend on client work.
So I built agents to handle it.
The Architecture
Meet Hank and Roofus, they each have their own role, focus, skills and knowledge.
Hank is my chief of staff. Scheduling, briefings, deal pipeline monitoring, meeting recaps, prospect research. He runs on cron jobs and pushes output to me over Telegram before I think to ask for it. Hank updates my project tracking tool, checks my emails, and forcefully keeps me on track
Roofus is my engineering partner. I use Roofus to explore new data pipeline builds. He' is building a relatively elite knowledge base of API documentation that ties in well with the DLThub skills (thanks DLT!), dbt best practices. Roofus builds and monitors pipelines for me as a POC, monitoring for failures and alerting me if there is a problem… fixing it independently when possible.
Both run on an open-source agent framework, Hermes. Each has its own Sould.md, its own tool access profile, its own cron schedule. They share memory through an outside memory system, more on this another time, iterating and improving on wiki's that centralize context across, but also specific to their own roles.
The choice that actually changed things is that the agents don't wait for me. They execute on a schedule. The morning briefing fires at 8 AM whether I'm awake or not. The pipeline review runs every Monday to start my week. Running and checking data pipes and models. The agents support my work, and nudge me in ways to keep me focused and moving forward
What's Working
One month in, here is what I love:
The daily cannabis digest is the one I'd miss most. Every morning I get a consolidated briefing covering federal policy, market-level changes across active states, and operator-specific news. Five minutes of reading instead of thirty minutes of tab-hopping. Call it ten hours a month I'm no longer spending in a browser.
The Monday pipeline review queries Linear for active deals, checks last-touch dates, flags the conversations that have gone quiet, and hands me a priority-ordered list. This is something I could do on my own but would be far less disciplined with, and would only be good as my memory.
Meeting recaps run after every call. Hank turns my notes into action items and risk flags in about two minutes. The bigger win is that "I'll remember to follow up" stopped being my actual follow-up system.
The one I didn't see coming is the multi-agent wiki work. Roofus researches and drafts, Hank formats and cross-references. Two agents editing the same vault, one producing and one checking the other. I went in skeptical of the whole multi-agent idea, and this is the thing that changed my mind.
Add it up and it's roughly 20 hours a month back. Adding my hourly rate to that provides their realized value.
What Broke, and What I Fixed
Nothing worked on the first try. Here's the honest version.
Memory was the hardest part by a wide margin. Figuring out what to persist versus what to regenerate took me longer than building either agent did. Store too much and the agent gets confused by stale context. Store too little and it loses the thread between sessions. There are a few approaches to this I have explored, I landed on a mix of solutions but this should be evaluated agent by agent
System prompt tuning turned out to be craft, not configuration. Hank's mode is "direct, flag problems bluntly, no hedging," and getting there without tipping into either robotic or sycophantic took a bunch of passes. Roofus is made to offer me options, could do A,B or C… pick one Ryan.
Tool access is where I'm most cautious, and probably where you should be too. Every capability you hand an agent is also an attack surface. Terminal access, git push rights, email reading, web extraction. Each one is fine on its own. Combined, they're real blast radius if an agent reads an adversarial web page and acts on it. For a solo operator the risk is acceptable. For a 50-person company running 50 agents, the math flips completely, which is why I'm building human-in-the-loop gating now, before something goes wrong rather than after.
Silent failure scared me more than loud failure. A few of my cron jobs depend on another job's output, and when an upstream job failed quietly, the downstream one happily produced garbage I might have acted on. So every job has a verification step now: did the thing it depends on finish, and finish cleanly? If not, stop and tell me.
Multi-agent collision is real, but it's fixable. Two agents writing to the same wiki repo will occasionally step on each other. The fix was boring and it worked. Explicit handoff rules, one agent produces and the other reviews, and they don't write to the same file at the same time.
The last lesson is the one I keep coming back to. The most useful thing I built into these agents was something I needed for myself. Hank has "shiny object defense" in his system prompt now. When a task starts pulling scope, he names the cost and asks whether the detour is worth it. I'm a solo operator who runs on curiosity, and so is he. That's not an accident. You're not really configuring software when you do this work, you're modeling a way of working, and the agent inherits your worst habits unless you tell it not to.
Where the Gaps Are
This is not an enterprise deployment, and I'm not going to pretend it is. The gaps, plainly:
My agents take instructions over Telegram. Convenient, yes. An authenticated, audited control plane, no.
Prompt injection is a live risk. If an agent ingests a web page, email, or document with adversarial instructions buried in it, it could act on them. Tool restrictions blunt this. They don't erase it.
Subagents spawn with their own tool access, and containing the blast radius is manual. I define per-subagent tool profiles by hand. There's no runtime isolation layer sitting underneath that.
Secrets are the part I'm actually happy with. Credentials live in environment variables via Doppler, never hardcoded, rotated on a schedule. Could be any keyvault service, totally enterprise ready.
Everything an agent does gets logged, but nothing is watching those logs in real time. If one went rogue I'd catch it on the next review cycle, if ever, not the moment it happened.
For a solo operator, that profile is fine. The upside is 20 hours a month and the worst realistic case is a bad email going out under my name. For a 50-person company running 50 agents it's a different conversation. Least privilege by default. Human gating on anything destructive.
Why This Matters for Operators
I would venture a guess that anyone reading this has been told to "go do agentic AI" by someone with a fancy title. I used my own business as a proof of concept here, the scale and complexity at enterprise level are different. The core ideas of defined agent roles, tight context, memory matters is very similar.
Building this took a week or so of trial and error, maybe ~$100 for the local server running my agents and minimal token costs to power my team. Your use cases are likely larger and more complex but the tools exist in the market to execute now. What you really need is good process hygiene, security discipline and focus on execution.
This works because I own my process. I know my process. I am empowered to integrate AI into my processes. This has firmed up my belief that AI projects fail when you approach them as technology work and not core business capabilities. This also means, while I believe you and your team could build dozens of agents to accelerate & improve core work to your business; you will need folks that have direct process ownership in the loop, on the train, whatever analogy you want here.
What's Next
A few things on the build list:
- Home management agent: partially built
- Additional cross agent workflows: I am being methodical and patient here but there is massive room
- Migrate wiki vaults to a knowledge graph: ambitious but scalable
- Exploring additional long term memory solutions (Honcho, Hindsight)
None of this is theoretical, Hank has it on my backlog already
If you're running a cannabis operator or a mid-market company and you're wondering what an agent could actually take off your plate, I'll give you a straight answer before I give you a proposal. Sometimes the honest answer is "not yet." I'll tell you that too.