Declared data loops
99
Scheduled jobs, declared with written rules for what each may read and write.
CASE STUDY · DIGITAL INFRASTRUCTURE
Casitel: 78 agent roles and 99 data loops, declared and built
An AI investment bank for digital infrastructure. Its agent roles and data loops are written down and built inside rules. People decide money, sends and deletes.
Our own platform
Casitel repository metrics
Source: Casitel repository, GitHub, 28 Oct 2026
Agent roles
78
Each role is defined in its own written file.
Pull requests merged
1,224
Changes merge only after automated checks pass.
Automated test files
488
Test files in the repository.
Database migrations
69
Schema changes tracked in version control.
Figures stated by Casitel
As stated by Casitel, Oct 2026
20,000+
Industry contacts
400+
Custom datasets
100+
Projects developed by the team
100+
Countries screened
The challenge
Casitel finds the site, the capital and the tenant for data centre and power projects. The work is high stakes and detail heavy.
A proven advisory service had to scale beyond the people who sold it. More context, more drafts and more checks could not mean more chaos.
The team needed one engine that gathers and scores inside clear rules. People stay in charge of commitments.
The approach
Four steps, in order. Each one sets a limit before the next begins.
- STEP 1 OF 4
Context before code
Map the decisions, data sources and hand-offs first. No agent is built until the business story is written down.
- STEP 2 OF 4
Write the job down
Each loop gets a brief: what it may read, what it may write and when it runs. Agent roles get the same treatment in their own files.
- STEP 3 OF 4
Agents do the work inside rules
Agents are set up to gather, score and draft. They may not send money, messages or deletes. Those stay with people.
- STEP 4 OF 4
People approve what matters
People read the drafts. People sign off on anything that binds the firm or a counterparty.
What was built
The design links sources, scheduled jobs, agent roles, review and people to outputs. The diagram shows the declared design. It is illustrative, and it does not show every integration or say that every loop is live.
Sources
Market, project and counterparty context
Loops
99 declared jobs
Agents
78 written roles
Review gate
Reviewer agent and CI
Decision
People approve binds
Outputs
Drafts, scores and reports
Feedback returns to sources (illustrative).
How the build runs
Software is built with the same method Mashroo sells. Cursor coding agents work from written briefs. A reviewer agent checks changes in a clean copy of the code. Nothing merges until automated checks pass.
- STEP 1
Written brief
Scope and rules before any code change.
- STEP 2
Cursor coding agent
Implements the brief in the application codebase.
- STEP 3
Reviewer agent
Checks changes in a clean copy of the code.
- STEP 4
Automated checks
Automated checks must pass before a change merges.
- STEP 5
Merge
A change merges only when automated checks pass.
What the numbers show
99 data loops are declared. Each loop states what it may read and write. The count shows what is written down and built. It does not show what is running.
78 agent roles are defined. Each role sits in its own written file.
1,224 pull requests are merged and 488 test files are in the repository. Changes merge only after automated checks pass. The first three changes the reviewer agent checked each had real defects, and it caught them (26 Oct 2026).
69 database migrations are tracked in version control, like application code.
Source: Casitel repository, GitHub, 28 Oct 2026
What this means for your business
Flow
Loops and schedules mirror Flow: repeatable operations with one owner and a written scope.
Learn
Role files and briefs mirror Learn: the method is written down so your team can read it and reuse it.
Reach
Casitel drafts sit inside written rules, and people choose what sends. Reach follows the same split between drafting and sending.
Build
Casitel software is built the way Mashroo builds for you: written briefs, coding agents, a reviewer agent and CI.
Studio
Advisory packs and narratives use the same context-first discipline as Studio.
Custom
Digital infrastructure advisory is bespoke by nature. Custom covers work that does not fit a standard product.
Want this discipline on your codebase?
Start with one measured loop or one product lane. We agree the measure first, then build inside rules you can audit.