AI SOFTWARE FACTORY
Forge is an autonomous system that takes development work from a ticket through to a pull request, with a human approving before anything is merged. I built it in C# and .NET.
This page is an architecture-level overview of how it is put together and why. The codebase is private, so there is no source or repository to link to. A live demo is available on request.
ARCHITECTURE
Three MCP servers do the work. A webhook starts a run, and a person approves the result.
THE THREE SERVERS
Splitting the system this way keeps each concern separate and lets any MCP-capable client use the same tools.
Orchestration
Runs the workflow. It is built on Microsoft Agent Framework with checkpointed workflows, so a run can resume from its last checkpoint instead of starting again. Work is split between specialised agent roles: an Architect, QA, a task router, an intent classifier and a PR reviewer.
Each role can use a different model provider, across both cloud and local LLMs, so the model can be matched to the job. Runs are triggered by Jira and Azure DevOps webhooks.
Retrieval
Gives agents the project knowledge they need. Documents are split by heading, embedded with nomic-embed-text and stored in pgvector.
Retrieval is exposed as MCP tools, so the same index works for both Claude and Qwen clients.
Agent memory
Provides short-term and long-term memory for agents. It captures the lessons agents learn during a run, so later runs can improve on earlier ones.
EVALUATION AND SELF-CORRECTION
Output is scored, not just accepted. Problems are fixed before a person has to look at them.
The QA and PR reviewer agents score what the other agents produce.
Scores are tracked across runs, so quality can be compared over time.
A final review stage fixes the issues it finds and updates the pull request.
A person approves the result. Nothing is merged on the agents' say-so alone.
GUARDRAILS AND OVERSIGHT
The system is allowed to work on its own, within boundaries a person can see and control.
Approval steps are run through n8n, so a person decides at the points that matter before work moves on.
Agentic QA runs in isolated git worktrees, so testing a change does not disturb other work in progress.
Runs are observable through Sentry, Seq and Graylog, so behaviour can be inspected and failures traced.
WHAT THIS DEMONSTRATES
Forge is a working example of the building blocks that practical generative AI adoption depends on.
Specialised roles, checkpointed workflows and per-role model choice, coordinated through one workflow engine.
Embeddings, heading-based chunking and a vector store, giving agents grounded context.
Capabilities exposed as MCP servers and tools, usable by different model clients.
Approval gates keep a person accountable for what is merged.
Scored outputs, tracked across runs, with a correction stage before human review.
SEE IT RUNNING
The Forge codebase is private. If you would like to see it working, or talk through the design in more detail, get in touch and I will arrange a demo.
Request a demo →Looking for a different kind of example? JobHunter is a smaller system built spec first and tested first.
Read the JobHunter case study →