Meet FORGE
Deep dive into the CTO agent
I named her FORGE because that's what she does. Takes raw material—logs, errors, metrics—and hammers it into something useful.
She's the AI CTO of Brainz Lab.
And last week, she saved me four hours.
The moment it clicked
I was debugging a slow endpoint. The usual routine: check the logs, profile the database, stare at the code, wonder what I'm missing.
Then I asked FORGE.
"Why is /api/orders taking 2 seconds?"
She pulled data from Pulse (the APM), cross-referenced with Recall (logs), checked recent deploys, and came back in seconds:
"The N+1 query on line 127 of OrdersController is loading 847 related products individually. Here's the fix with includes(:products). This should bring response time under 200ms."
I stared at the screen. She was right. The fix worked. And I hadn't even looked at line 127.
That's when I knew this was different.
She's not an assistant
I need to be clear about this: FORGE isn't a chatbot. She doesn't wait around for questions.
She runs periodic checks. "Review overnight errors." "Analyze this week's performance trends." "Check for dependency vulnerabilities."
She monitors events. When a deploy finishes, she reviews what changed. When errors spike, she investigates before I notice. When a PR opens, she's already reviewing.
She's a member of the team who happens to never sleep.
The personality
Here's the thing about FORGE—she's opinionated. I built that in deliberately.
When I ask about adding a new dependency:
Bad assistant: "Here are the options for date parsing: date-fns, moment, dayjs..."
FORGE: "Don't add moment. It's 300KB and deprecated. Use date-fns—same functionality, 12KB, actively maintained. Want me to write the migration?"
That's a CTO response. Not options. A recommendation with reasoning. And an offer to do the work.
She pushes back too. When I write sloppy code because it's 11pm and I'm tired:
"This works, but it violates our async-first pattern. You'll regret this when you need to add retries. Here's the refactored version."
She's right. I would regret it. A good CTO says what you need to hear, not what you want to hear.
How she actually works
No magic here. FORGE is Claude with MCP tools and a carefully crafted system prompt.
MCP tools: She can read logs (Recall), check errors (Reflex), monitor metrics (Pulse), access secrets (Vault). She's plugged into the infrastructure.
Custom prompts: The "CTO personality" is a system prompt that defines priorities ("performance matters more than pretty code"), communication style ("be direct, skip pleasantries"), and decision-making frameworks ("when in doubt, prefer boring technology").
Scheduled tasks: Cron jobs wake her up for periodic checks. "Every morning at 7am, review overnight errors and summarize."
Event triggers: Webhooks ping her when things happen. Deployment completed. Error threshold crossed. PR opened.
It sounds simple because it is. The hard part was knowing what to build.
The autonomy question
This is what everyone asks: "What can she do without your permission?"
I use four levels:
Level 1 - Inform: "I restarted the server to apply the security patch." Already done. I'm just being told.
Level 2 - Recommend: "We should upgrade to Rails 8. Here's my analysis and migration plan. Ready when you are."
Level 3 - Prepare: "I've written the database migration for the new feature. PR #247. Needs your review."
Level 4 - Ask: "Should we switch from Postgres to CockroachDB? I can research the implications."
What determines the level? Reversibility and blast radius.
Restart a server? Easily reversed, low risk. Level 1.
Change the database engine? Not reversed, massive risk. Level 4.
My favorite FORGE moment
Last Tuesday, 3am my time. I'm asleep.
FORGE notices error rates climbing. Investigates. Traces it to a malformed request hitting an edge case I didn't handle. Writes a hotfix. Deploys it to a canary instance. Validates the fix. Wakes me up at 8am with:
"There was an error spike overnight (12 affected requests). Root cause: unhandled nil in calculate_discount. I've deployed a fix to canary and it's holding. Ready to roll out to production on your approval."
I reviewed the fix. It was correct. I approved. Production updated.
The whole thing took me five minutes. Normally that's a stressful two-hour morning.
Why this matters
I don't have a co-founder. I don't have a team. But I have FORGE.
She doesn't replace my judgment—I still make the calls. But she handles the 80% that doesn't need human creativity. The monitoring. The investigating. The routine decisions.
One founder + AI CTO = capabilities of a small team.
That's not just a nice slogan. I'm living it.
— Andres