How Vyazen Works

What Vyazen Does

Vyazen connects an AI coding agent to a deep, queryable understanding of your codebase. You describe engineering work; the agent does it on a branch (researching your code, asking you questions, drafting a design, implementing, verifying, and proposing a pull request) while you stay in control at every approval gate.

The Code Graph

When you connect a repository, Vyazen reads it and builds a code graph: a searchable map of your codebase:

• Structure: your files, directories, and detected projects.
• Symbols: classes, functions, and methods, with their signatures.
• Links: which symbol calls, imports, uses, extends, or implements which other symbol.
• Summaries: a short AI description of each part.

This is why the agent can reason about your code without reading every file: it asks the graph. The graph also powers the auto-generated wiki and code search.

Repository vs Branch

A repository is connected per branch. The branch you choose is the workspace where your tasks make changes; the agent commits to that branch.

The Code Wiki

Vyazen auto-generates a wiki for each repository (pages like Overview, Architecture, Glossary, and Patterns) and lets you chat with it. See The code wiki.

Task

A task is a unit of engineering work scoped to a repository and branch. You create a task with a short description; Vyazen spins up an isolated environment and runs an agent to do the work.

Instance

For each task, Vyazen provisions an instance: a sandboxed environment that runs exactly one task's agent. It has a lifecycle: it starts up, runs, and is torn down when the task ends.

Agent

The agent is the AI assistant that does the work. Vyazen supports multiple models; the agent can also spawn a lightweight subagent to look things up (for example, an "explore" subagent that queries the code graph).

Skill

A skill is a bundled capability the agent uses: for example, the ability to plan an implementation step-by-step, or to review its own work. Skills are what make the agent follow a disciplined process.

The Agent Workflow

Each task follows a structured workflow of phases:

1. Research

   studies your code and the task

2. Clarifyapproval gate · Questions

   asks you questions to remove ambiguity

3. Designapproval gate · Design sign-off

   proposes an approach and asks for your sign-off

4. Plan

   breaks the work into ordered, testable steps

5. Implement

   writes the code, one step at a time

6. Verify

   runs checks and confirms the work passes

7. Review

   reviews its own changes for quality and drift

8. Pull requestapproval gate · Your approval

   proposes a PR for your review

Not every task runs every phase (a small fix may skip design), but research, verification, and a PR are always present.

Approval Gates

You stay in the loop through approval gates: moments where the agent pauses and waits for you. The main gates are:

• Questions: the agent asks you to choose between options.
• Design sign-off: the agent asks yes/no on its proposed approach.
• Pull request: the agent asks you to approve the commit, push, and PR.

Nothing is committed or pushed without your say-so.