Agent
Graph
A graph engineering workspace for discussing, dispatching, and observing multi-agent work without leaving Visual Studio Code.
release 0.1.6 · MCP control plane · Docker workers
The runtime gap
Powerful models can reason about code. Engineering still needs explicit ownership of issues, routes, permissions, evidence, retries, and completion.
Human direction becomes a governed graph
The Manager Agent translates a request into an issue and chooses the required specialists. MCP carries the contract. The Dispatcher executes one isolated worker at a time and decides what may enter the queue.
Every required agent already has a place in the queue. Dynamic handoffs add useful work; they do not decide whether mandatory specialists get a turn.
A dynamic graph without uncontrolled execution
Workers may propose the next agent—or several agents—with a focused prompt for each. The Dispatcher treats every proposal as untrusted input.
Intersect the graph
The next worker must be selected for the issue and authorized by the current agent’s saved allow-list.
Merge work before it starts
New prompts for a waiting agent are consolidated without moving its queue position.
Append deliberate loops
If an agent is working or has finished, a new valid assignment creates another visit instead of rewriting history.
Bound every cycle
Step, per-agent visit, and final rework limits stop the graph deterministically.
One system, four visible surfaces
The extension keeps conversation, tickets, team configuration, and execution state close to the code while separating each concern.
Chat
Talk to the Manager Agent at project scope or system scope. Human requests, Manager outlook, agent progress, and useful outcomes stay readable in one editor tab.
Issues
Inspect newest-first tickets with description, graph queue, live activity, agent reports, final verdict, and expandable history.
Agents
Manage the roster, model, reasoning depth, role, and allowed routes. Protected role instructions are restored before every new worker run.
Graph
See the effective team and routes. Solid green, yellow, and red nodes show idle, working, and error state while the Dispatcher continues valid pipeline work.
Clear ownership at every layer
Graph agents do engineering work. They do not need to understand the extension, mutate orchestration state, or call each other.
Human
Defines intent, constraints, and acceptance.
Manager
Discusses, creates issues, and selects the team.
MCP
Exposes the typed control-plane contract.
Dispatcher
Owns the queue, validation, execution, and recovery.
Workers
Reason, use tools, change code, and report structured results.
Context should support work, not replace it
The current assignment is always first. Direct check-ins receive no filesystem references. Engineering work can consult only the evidence it needs: the project workspace, issue discussion, related tickets, chat history, and private same-agent memory.
Durable evidence
Chat is for people. Issue JSONL is for collaborating agents. Raw logs are for machine audit and diagnosis.
Codex execution inside a controlled container
Each graph turn runs through a pinned Codex CLI in a versioned Docker image. Project workspaces and private agent notes are writable; issue and chat references are read-only; runtime credentials and MCP controls stay outside the worker.
Provider
Use an OpenAI-compatible vLLM Responses endpoint or isolated Codex subscription login shared only within Agent Graph.
Network
Choose Docker bridge or host networking without rewriting the configured inference URL. Readiness verifies a real response through the selected route.
Recovery
Detached workers and durable queues let the extension reconnect to active execution after an extension-host reload.
Start with one conversation
Set up the provider once, then ask the Manager Agent to explain the current team or optimize the graph for a project.
Install Agent Graph
Install the VSIX from Open VSX, then open the extension from the Visual Studio Code Activity Bar.
Connect inference
Use the default isolated Codex subscription login or configure a vLLM URL and API key.
Build the worker
Save Docker settings to build and verify the exact versioned runtime image.
Ask the Manager
Open system Chat to inspect the roster, then add a project and dispatch engineering work.
Make the work visible. Keep execution deterministic.
Agent Graph gives specialized models room to reason while software owns the queue, graph, permissions, evidence, and stopping conditions.
Agent Graph 0.1.6