Building a Lean AI Team with Brennan Dunn (Founder of RightMessage)

Brennan runs RightMessage solo. No full-time hires, just a rotating bench of freelancers and an AI stack doing the work of what used to be a team. RightMessage is growing faster than it did with a bigger team, and Brennan is stuck on a question a lot of founders are facing, not "how do I scale my team," but "do I even need to?"

In this session, Craig Hewitt, Founder of Castos, sits down with Brennan to dig into what a genuinely lean, AI-run company looks like day to day, along with sharing his own experience running Castos, which roles AI has completely absorbed vs. where humans are still needed.

If you're a founder wondering whether your next move is a hire or a better AI workflow, this is the conversation to sit in on. 

Come ready with your questions!

Brennan raised money for RightMessage in 2018 and built out a full team, which in hindsight pointed the company in the wrong direction. Over time he realized he didn't enjoy managing people or the "job" that running a team created for him — a calendar full of one-to-ones instead of building. Two influences pushed him toward going solo: Laura Roeder's blunt "we're professionals, not a family" philosophy, and Jesse Hanley from Bento, who runs a full email marketing platform essentially by himself. 

As agentic AI tools (Claude Code, and later custom setups) got capable enough to outperform his junior developer and tier-one support contractor combined for a fraction of the cost, he let go of his last team member around the end of 2024 and has been running RightMessage solo since.

Where AI is easy vs. hard

Both Brennan and Craig agreed marketing and most of product are the easy wins — content, blog posts, features — because you control the whole environment. Support and customer success are trickier. For Craig, that's because a lot of Castos's support volume comes through a WordPress plugin he doesn't fully control. For Brennan, support has actually become the easy part (he estimates he personally handles only about 5% of tickets now), while success and sales have become his hardest, most time-consuming work.

How Brennan runs support almost entirely on AI

Every inbound ticket gets classified as either a "success bug" (a strategy/how-to question) or a "product bug" (something's actually broken). His AI has access to the full support inbox, ticket history, and the codebase, so it can recognize patterns ("three other people asked this same thing") and — when appropriate — actually implement a fix, open a pull request, and notify him to approve and deploy. Every ticket is treated as a signal to reduce future tickets, not just a fire to put out: recurring questions get turned into newsletter or video content, or trigger onboarding/UI improvements. His first-response email is openly AI-generated and transparent about it (it references help docs and searches the codebase for exact button/screen guidance). A separate admin MCP gives his own account roughly 30 extra tools other team members don't have — things like extending a trial or pulling account configuration — so common asks (like "my trial expired, can I get more time") are resolved automatically via a simple SOP.

The daily infrastructure

Brennan walked through his actual setup:

• A VPS running a Hermes-based agent reachable through a Matrix chat client on his phone, giving him a way to direct work even away from his laptop (including merging PRs, which triggers deploys).
• That agent has access to a long list of MCPs — Stripe, PostHog, Bento, Sentry, GitHub, and his internal admin tools — plus OpenCode/OMP as a harness for autonomous, greenfield feature work: he can hand it a Linear ticket from his phone and it will build the feature and open a PR.
• A markdown-based "vault" that mirrors his help docs, marketing site, sprint plans, and even personal goals, which the agent uses as long-term memory and pulls from to draft things like investor updates.
• An automatically generated daily log (built from email, Git activity, etc.) that makes writing investor updates or recalling "what happened the last two weeks" trivial.
• A walking email-triage routine: he talks to his phone, the agent archives junk, summarizes threads, drafts replies, and he verbally edits and approves — getting him to inbox zero without sitting at a desk.
• A weekly SEO/content loop that reviews Search Console and Ahrefs data and proposes content updates to the marketing site via pull requests (currently on a Monday cron job — he flagged the lack of a staging environment as a real pain point here).
• Sales-call support: the agent researches prospects beforehand (including any prior support history), scores his performance after each call, and manages automated, considerate follow-up nudges — though he was candid that this "CRM" is really just markdown files and a JSONL feed and wouldn't scale to a second salesperson.
• A Sentry integration that detects new exceptions, attempts a fix, opens a PR, and pings him to approve — letting him fix production issues from his phone without ever opening his laptop.

A theme he kept returning to: the biggest mental shift was moving from him having to go talk to AI, to setting things up so AI proactively surfaces only what actually needs his attention — closer to managing a team that filters noise for you than to babysitting a chat window.

Open problems he's still working through

Brennan doesn't have a good answer yet for running multiple coding agents in parallel on different features — git worktrees separate the code fine, but he hasn't solved running multiple simultaneous instances of the app/dev server. Andrew shared how his team handles this: maintaining a fixed set of dev worktrees (4–5), each wired up as its own persistent app instance with its own local URL and Cloudflare tunnel, so several agents (and engineers) can run and test in parallel without needing to spin up new environments dynamically. Brennan liked the idea and plans to adopt it. He also admitted key-person risk is real — he keeps his laptop within 30 minutes of him at all times ("the nuclear football") for the rare cases (like a Cloudflare outage or server issue) that need hands-on intervention from a real terminal, since he's not yet comfortable running one from his phone.

Code review in an AI-written codebase

Brennan doesn't review every line of code anymore. He leans on an AI code-review tool (he's tried Code Rabbit, Greptile, and now Cubic) for a risk score on each PR, and has a custom "slop check" skill that runs before a PR goes up to flag opportunities to simplify or refactor. He was candid that he's lost some detailed knowledge of his own backend architecture as AI has taken over more of the writing — but bug counts are down, not up, and he's made peace with thinking of himself as a "product founder" rather than a "developer founder." Craig's team at Castos still has human engineers manually review every PR in addition to AI review tools. One other tactic mentioned: using Claude to write code and Codex/ChatGPT to review it, since a different model tends to catch more than the one that wrote the code.

Getting a whole team (not just the founder) to actually use AI

One founder asked how to push AI adoption across a team when they don't know each team member's job well enough to show them the shortcuts. Craig's approach at Castos: make it an explicit, non-negotiable expectation ("this is just how we work now — the first question for any task is how can AI do this with or for me"), paired with a weekly show-and-tell meeting where everyone is expected to bring something they automated or improved with AI that week. Social pressure did the rest — people who showed up empty-handed a couple times got the message. Both Brennan and Craig framed the goal not as replacing people, but as freeing them from repetitive work so they can grow into higher-value roles (Craig mentioned a support teammate who's on track to become a full-stack developer simply because there's much less manual support left to do).

Other tools mentioned

FernDesk came up as a help-desk tool Brennan is happy with — it connects to Help Scout and GitHub, notices when a merged PR should update a help doc, and proposes the diff automatically, which has kept his documentation from going stale without him having to enjoy the process of writing it.

Where Brennan sees this going

He was clear that he's not trying to build "the next Intercom" — he describes running RightMessage more like a neighborhood pizzeria: profitable, low-stress, with plenty of time for tennis and family. His bigger long-term concern is less about AI and more about how SaaS as a category might shift given how easy building software is becoming (he compared it to the "on-premise software" disruption cycle), but he's not losing sleep over it.