PUBLIC REPOSITORY · EXPERIMENTAL
Guarded Agent Pipeline
A multi-model development pipeline with checks and an explicit human boundary
A workflow that coordinates Codex, Claude Code, and DeepSeek with deterministic verification and human-controlled pushing, so model agreement is never mistaken for engineering correctness.
- TypeScript
- MCP
- PowerShell
- Git
EVIDENCE
Public material in progress
Real interface captures and reproducible material are still being prepared. This page does not substitute invented screenshots or unverified numbers.
SYSTEM MAP
How it works
- 01Constrain the task
- 02Model implementation
- 03Independent review
- 04Deterministic checks
- 05Human approval
The problem
Having several models review one another broadens the search for mistakes, but it does not guarantee correctness. Models can share blind spots or agree without ever running the tests. Development automation needs checks outside the models themselves.
Design
The pipeline separates implementation, independent review, deterministic verification, and human approval. External scripts run tests, inspect the working tree, and record artifacts. Models propose and explain changes. The final push remains an explicit human decision.
Current progress
- A limited, structured set of pipeline operations exposed through MCP.
- Auditable inputs and outputs saved for every stage.
- Explicit states for failure, recovery, and cancellation instead of guessing from chat history.
- A PowerShell recovery path that reduces dependence on one orchestration entry point.
The core question is not how to add another agent, but how to let several uncertain components cooperate inside deterministic guardrails.