Case study 03 · Team and operations

Multi-agent AI team

SEO · marketing · profile · product · tradingApproval gatesKill switches

I build Quantlys at night and on weekends, alongside a full-time job. The team that makes that possible is mostly AI agents, plus a few remote human teammates. This is how it's set up, where the approval lines sit, and what went wrong.

01 · The problem

One person, many jobs

The work spans Quantlys Conclave (the AI app builder), Tablerene (an AI phone host for restaurants), Quantlys Meeting (video meetings that write a PRD), a paper-trading system called QTP, and the marketing and SEO for all of them. There isn't enough of me to do each job by hand. And one all-purpose AI agent with no limits is a risk I won't take.

02 · The approach

Narrow lanes, clear gates

  • Split by role. Each agent owns one lane: SEO, marketing, profile, product or trading.
  • Gate what can't be undone. Agents do the work, but anything public, paid or sent waits for my approval. One agent saying I approved something doesn't count.
  • Keep a short list for myself. I delegate technical calls and keep four things: money, legal, live keys and pricing.
  • Same shape inside the product. Conclave runs a PM, Architect, Developer and QA crew through Draft, Peer-rank, Arbiter and Ship.
  • One branch for people and agents. AI coding lanes on two Macs and remote human teammates all ship to the same main.
  • Design before build. Features go through an Elephant-Goldfish design review: a concepts brief, then a design spec, then reviewers with no prior context read it cold, then approval.
  • Hard bars where money is involved. QTP runs on a paper account. A strategy has to clear a measurable go-live bar before it touches real money (for the B10 strategy, 30 to 50 trades with a profit factor above 1.3), and the system has per-strategy kill switches, a pause switch and a dead-man's switch. B10 hasn't met that bar yet, so it stays on paper.
  • The basics everywhere. Code in git, rollbacks that work, and no secrets in code.
03 · What happened

It ships, and it breaks

The setup ships real software. Tablerene's first version, with 24 pages, 55 routes and 23 database migrations, was built by directing Conclave with written prompts and reviewing the pull requests, not by hand-coding. Quantlys Meeting was built with Conclave too. You can see them on Built with Quantlys.

It also broke in ways I caused. On November 4, I deleted a credential in QTP's n8n workflows without checking what depended on it, and 10 workflows broke for about 10 minutes. AI lanes sat idle until I prompted them. Teammates shipped to main without posting first. On Tablerene, migrations weren't applied to the live database. And approval gates create a queue: owner-only tasks pile up and can block a launch more than code does.

The biggest cost is breadth. Running this many products at once puts every date at risk.

04 · Lessons

What running it taught me

  • Narrow lanes beat one general agent.
  • Gate the irreversible: anything public, paid or sent, plus money and keys.
  • Approvals are a queue. Clear them on a schedule, or they become the bottleneck.
  • Check dependencies before you delete anything.
  • An idle agent is a coordination bug. Give every lane its next task, not a wait.
  • Fewer bets, taken further.
05 · For another team

Before your second agent

Write down the lanes and the approval list before you add a second agent. Hold people and agents to the same gates. Block out a daily slot to clear approvals. And judge the team by what reached users, not by how much it produced.

About the author

Santosh Adari

Santosh Adari is Associate Director, IT Data Engineering at Gilead Sciences and the founder of Quantlys, which he builds on nights and weekends from Apex, North Carolina. Quantlys is a native app for Apple Silicon Macs; inside it, Conclave plans an app with a council of AI models, then builds it and deploys it to your own GitHub, Vercel and Supabase accounts.

About Santosh Adari · All case studies · Built with Quantlys · LinkedIn