AI Agents · Orchestration
Lang-Agent
A multi-agent development system where a supervisor decomposes a task and dispatches specialist agents into Docker sandboxes, streaming every thought and tool call to a live dashboard.
Organisation
Independent Project
Role
Solo — architecture, agents, infrastructure
Timeline
2026
Stack
Overview
Lang-Agent is a team of AI agents that builds software. A supervisor named Igris reads the task, decomposes it, and dispatches specialists — Beru on backend, Kira on frontend and design — either in parallel or in sequence, then aggregates what comes back. The whole run streams to a React dashboard: thinking, tool calls, files written, and any point where the supervisor decides it needs a human.
The interesting constraint was that runs are long. A real task can take hours, so the system is built so you can submit work, close the browser, and read the full replay in the morning.
Key Features
- Supervisor orchestration on LangGraph, using langgraph-supervisor with handoff tools per agent and the Send API to dispatch backend and frontend work simultaneously rather than serially.
- Skills as files, not prompts — each agent owns a directory of SKILL.md manifests (decompose-task, delegate-to-agent, risk-assessment, synthesize-results, observe-progress for the supervisor alone) that a loader concatenates into its system prompt, so capabilities are added by writing a file.
- Docker-sandboxed execution — agents that write and run code do so in constrained containers with memory and CPU limits and no network, keeping generated code away from the host.
- Durable background runs — a SQLite task queue and a separate worker process, with every action appended to a persistent event log (agent, type, payload, human-readable summary) so a completed run can be replayed in full.
- Human-in-the-loop by the supervisor’s judgement — Igris calls interrupt() itself when an action looks risky, rather than gating every step, and the dashboard surfaces the question for an answer.
- MCP tool integration via langchain-mcp-adapters, giving the frontend agent browser and design tooling alongside its own file operations.
- Token-level streaming from astream_events() through an Express bridge to the browser over SSE, plus a Python test suite covering models, tools, events and the streaming path.
Impact
Most agent demos work in one shot and lose everything when the tab closes. Lang-Agent treats a run as durable state: a queue, an event log, a replay. Making skills declarative files rather than hard-coded prompts is what turns “an agent that codes” into something you can actually extend.