For some time I was not happy with how Claude Code (and other coding agents) behave. They often start coding too early, skip proper requirements, or declare the work finished when it is still incomplete. So I decided to create something that puts real structure around the agent.
The result is this repository:
https://github.com/tomas-rampas/claude-agentic-framework
It is not a new AI tool. It is a configuration layer that sits on top of Claude Code CLI and tries to make the agent more disciplined.
The core idea
I designed a clear pipeline:
Spec → Build → Review
Everything starts with the /delegate command. If there is no specification yet, the system automatically runs /spec. In the specification phase the agent asks questions one by one, then scores its own understanding. Only when I approve the status, the work can continue to the build stage.
During /build the system first maps every requirement to concrete files. Then it starts implementing. After that comes the Review & Fix Loop. This loop can run maximum three times. If the reviewers are not happy, the agent has to fix the issues and go through the review again. Only when the final status becomes “built”, the process is considered done.
There is also a Quality Bar that shows the progress (from red to green) and a per-todo BDD style loop.
Two important mechanisms
I added a Review Chain with three different roles:
-
spec-compliance-reviewer— checks every requirement one by one -
code-review-gatekeeper— looks at code quality before commit -
peer-review-critic— final independent review (this is the last gate)
And a Self-Scoring Loop that the agent uses when it needs to improve its own non-code output. It follows five steps: Rubric → Score → Name Weak points → Rewrite → Rescore. It continues until the improvement becomes very small.
At the bottom of the process there is an Enforced Stop Gate. Two PowerShell hooks make sure the agent cannot just end the session if the peer review is not satisfied.
What is inside the framework
Currently it contains:
- Around 20 specialized agents (language experts + roles like product-owner, system-architect, security-specialist, etc.)
- Real enforcement hooks written in PowerShell
- Commands such as
/delegate,/spec,/build,/review-spec - An anti-drift system that tries to keep the agent registry, documentation and files consistent
- Integration with several MCP servers
You still need Claude Code CLI. This project only configures and constrains it.
Why I made it
I wanted less “hope the agent does the right thing” and more real process. The agent should not decide by itself when the work is finished. There are hard limits and gates.
Of course the quality still depends a lot on how good the individual agent definitions are. But at least the structure is there.
If you use Claude Code and sometimes feel that the agent drifts or finishes too early, you can try this setup. Installation instructions are in the repository.
I’m still improving it, so feedback is welcome.
Top comments (0)