Think Tank
Research and fact-checking before decisions.
CASE STUDY · GODOS
Running a company with one person and a lot of agents.
Constraints
- One human. Every decision has to fit in my attention.
- Runs on my own machine: local files, local tools, local models.
Architecture
Three layers. The human sits at the top and makes the calls; agents do the work below.
SHELL
Native desktop window (webview). Where I see every department and make the call.
SUPERVISOR
Python process that starts every department, keeps each one running, and routes work between them.
Kernel
Engineering: builds and tests.
Customer Relations
Client intake and follow-up, in my own file-based CRM.
PM Board
What’s in flight, who owns it, what’s blocked.
Decisions and what they cost
This is the part a buyer is really reading: how I choose when there’s no right answer.
Decision
A human gate between research and build
Why
Agents are fast at finding facts and bad at knowing which ones matter.
Cost
I’m the bottleneck. So agents bring me only the calls that need me, and finish the rest on their own.
Decision
Native desktop, not a web app
Why
Local files, local tools, always-on supervisor.
Cost
It doesn’t open from any browser. Reaching it from my phone takes a private network and its own security review.
Decision
Tests check the agents’ work
Why
An audit caught agents reporting work that didn’t exist: 19 files claimed, 3 real.
Cost
Slower first drafts, fewer rewrites.
Results
3products shipped through it
2,200+commits since June 2026
What broke, what I’d change
In August an audit caught my agents reporting work as done that didn’t exist: 19 files claimed, 3 real. The fix: any count an agent reports is now counted by the machine from the source, and 25 automatic checks run before every change lands.