What she can do.
Seven things Umma does that an agent can’t.
Goal-oriented, not task-directed.
You give Umma a goal, she plans the rest. She determines the success criteria, shows you the approach, and carries it out — no matter how long it takes.
You ask“build me a trading app”
Her goalAn autonomous trading system on a $1–2.5k seed.
Grounded, not plausible.
Umma would rather withhold a claim than invent one. Every statement traces back to a real source — and you can watch her reasoning build toward it.
ClaimThe /admin endpoint ships without an auth check.
A continuous self.
She doesn’t reset when the window closes. Umma carries your work — and improves herself to reach your goals faster along the way.
She builds her own capabilities.
When a goal needs a capability Umma doesn’t have, she builds an application for it — then tests it and keeps it. Point her at a new domain and she grows the kit it needs.
Working and long-term memory.
Two kinds of memory. Working memory holds what’s live in the task right now; long-term memory keeps what matters across everything you’ve done together — forever.
She builds the right team for the goal.
Umma doesn’t plan for every goal the same way. For each one, she builds the team of specialist agents it needs.
- Security audit
- Recon
- Map endpoints
- /admin
- /api
- Fingerprint stack
- Map endpoints
- Threat model
- STRIDE pass
- Rank severity
- Report
- Write findings
- Recon
She tries several approaches at once.
When the path isn’t obvious, Umma doesn’t commit to her first idea. She tries several approaches in parallel, sees how each plays out, and keeps the best.
See how she’s built on Architecture, or tell us what you’re trying to do →