BcS
Co-founder, The Smith Syndicate

BrianCharlesSmith

I design agentic systems that ship to production, and the guardrails that let a company trust them.

Day job
Security Product Engineer, Enigma Networks
Building
Cruina, Omniira and Kiryn
Writing
Three papers on agent architecture

01Practice

How I run a team of agents.

Most of the code I ship is written by Claude Code sessions running side by side, which moves the hard part of the job upstream: knowing what to build, saying it clearly enough that it can be checked, running the team that builds it, and deciding where that team is allowed to go. Those four are how I work, and keeping an eye on the team while it works is the reason I built Kiryn.

  1. Define done

    Plans and acceptance criteria

    An agent builds exactly as well as the plan it was handed, so every piece of work starts as a written plan: the goal, the guardrails, what's out of scope, what success looks like, and acceptance criteria that each name the command that proves them.

    Before the build starts, a fresh agent that has never seen the conversation attacks the plan, looking for every way it could be built wrong while still passing what's written, because ambiguity is cheap to fix in a plan and expensive to fix in a build. Then each check runs against the untouched code, and any criterion that already passes gets thrown out, since it could never have failed.

  2. Lead the team

    Delegation and review

    Running six agents at once is management work. Every one of them needs an assignment it can't misread and the context to carry it out, reviewers work through their own slices of the change in parallel, and every build gets an adversarial review from an agent that didn't write it.

    The culture is written down: commit format, branch names, conventions and the reasons behind them live in instruction files that every session inherits and follows without being asked. The tooling asks me for very little, and when it does ask it matters, because a system that prompts for everything trains you to click yes and the prompt that mattered slides through with the rest.

  3. Think like a founder

    Product and vision

    Agents will build whatever you describe, which makes deciding what deserves to exist the scarce skill. Taking Cruina and Omniira from an idea to production taught me to start from the person using the thing and to keep a picture of where the product should be a year out, so this week's work points somewhere. It also taught me to know where the money comes from and what each customer costs to serve.

    There's room to be creative too. Omniira began as an experiment in whether AI characters could hold real memories, and it grew a government and an economy because the vision kept outgrowing the prototype.

  4. Draw the lines

    Permissions and the production boundary

    Some of my rules are culture, which can bend for a turn when I say so, and some are floor, enforced in code where no instruction reaches. A session can work across every repository and ship to stage on its own, and it can't read a secret or touch production even if I tell it to halfway through a session.

    WORKSPACESTAGESECRETSPRODUCTION

    Human only

    A hook on every shell command blocks anything aimed at production and fails closed, and permission I grant mid-session doesn't change that. When a task needs prod, the agent hands me the exact command with placeholders and carries on with everything else.

02Work

Three systems, all in production or close to it.

All three are ventures of The Smith Syndicate, the studio I run with my brother. Kiryn is also the tool I use to keep track of the sessions that build the other two, and all three were made the way the section above describes.

LiveCo-founder, engineering

Cruina

Private family reunion and event planning

181
API operations
1
codebase for iOS, Android and web
0
ways to join a family without an invite

A private home where a family organizes reunions, keeps a member directory and a family tree, takes registrations and collects money. Payments go straight to whoever is running the event through Stripe Connect, so nothing sits in escrow and any member can be the one collecting.

It's a multi-tenant SaaS, and the whole product rests on one invariant: every query against a tenant table takes the tenant from the verified session and never from the request. The API is Hono and TypeScript on Cloud Run with Terraform behind it on GCP, and one Expo codebase serves iOS, Android and the web.

cruina.com
LiveCreator

Omniira

A persistent world run by autonomous AI characters

29
named characters with memory
3
model tiers behind one queue
175
game days of logged behavior

Omniira is an ARC, my name for an autonomous realm: a multiplayer fantasy world where the people you meet are language-model agents who remember you, decide for themselves what's worth remembering, pass gossip along with a record of who said it, and write a diary entry at the end of every game day.

Above them sits a government, where four LLM ministers read the live economy and propose tax changes and warehouse releases, and a regent enacts at most one decree per cycle. Inference is split across three model tiers behind a decision queue, so the cheap calls stay cheap and the expensive ones are rationed.

omniira.ai
Coming soonCreator

kiryn

A console for running a fleet of Claude Code sessions

1
credential, and it can prove it
0
lines of code it writes itself

Once you run six agent sessions at once, the hard part is knowing which one is stuck waiting on you. Kiryn reads Claude Code's own transcripts and shows every session on the machine: its repo and branch, the tool it's running, how full its context is, and whether it needs an answer.

You drive it by voice, with a latch that decides where your words go and a voiceprint that refuses anyone else in the room. Kiryn opens terminals and carries keystrokes between them, but it never writes code itself and holds exactly one credential, which matters for a tool that's headed for open source.

kiryn.dev

04Path

Chef, sergeant, restaurant owner, engineer.

I came to software late and from a long way off. A kitchen line and a police sergeant's desk turn out to be good training for running autonomous agents, since both are about writing procedure that holds up when you aren't standing there, and knowing exactly who's allowed to make which call.

  1. 2008

    Culinary school, then the line

    Johnson & Wales University · Philadelphia

    Earned a BS in Culinary Nutrition, magna cum laude, then worked up to sous chef in the Greater Philadelphia area.

  2. 2014

    Philadelphia Police Department

    Philadelphia, PA

    Served eight years as a police officer before being promoted to sergeant in 2022.

  3. 2023

    Co-owner and executive chef

    Dawghouse Grille · Elmer, NJ

    Left the department to help start a food truck and a brewery and cidery in South Jersey.

  4. 2024

    Software engineer

    Self-taught

    Taught myself to code and built my first site by hand from a YouTube tutorial, which this page replaced.

  5. 2025

    Co-founded The Smith Syndicate

    With my brother Josh

    We build privacy-conscious software as an independent studio, and Cruina, Omniira and Kiryn are ours.

  6. Now

    Security Product Engineer

    Enigma Networks

    I build security products by day and agent-built ventures the rest of the time.

05Contact

If your team is working out how to put agents into production, let's talk.

brian [at] thesmithsyndicate.com