I Built an Open-Source Control Layer for AI Agents
AI coding agents are becoming increasingly capable.
But there is another problem:
What happens when you want multiple agents to work together?
One agent plans.
Another implements.
Another reviews.
Another tests.
Another decides whether the work should be accepted.
Without an orchestration layer, these agents can quickly become isolated processes with fragmented context.
That was the problem I wanted to explore.
So I built Orchestrator.
The idea
Orchestrator provides a shared environment where humans and AI agents can collaborate on software projects.
Instead of thinking:
«Agent A + Agent B + Agent C»
the architecture treats the system more like:
Human + Agents + Shared State + Tasks + Memory + Verification + Governance
The execution pipeline
A task can move through multiple stages:
Plan → Implement → Critique → Test → Arbitrate/Merge → Review → Record
Each stage produces persistent execution information rather than simply disappearing after the model responds.
Shared project rooms
Agents can operate inside project rooms containing:
- Tasks
- Messages
- Memory
- Runs
- Execution history
- Project state
This provides a common context for the team.
Agent integrations
The architecture supports adapters for different coding-agent workflows, including Codex, Claude Code, OpenCode and Cline, alongside a generic adapter approach.
The goal is to keep the orchestration layer independent from any single agent.
MCP
The project also exposes orchestration functionality through MCP.
Agents can interact with concepts such as:
- Rooms
- Tasks
- Memory
- Runs
- Replay
- Ledger verification
Safety and governance
Multi-agent execution introduces another challenge: giving agents useful capabilities without turning the system into an uncontrolled automation layer.
The project therefore includes controls around subprocess execution, browser actions, plugins, secrets, path traversal, injection filtering and execution budgets.
What isn't finished
This is important.
I don't consider the current implementation production-ready.
Areas such as stronger multi-tenancy, quotas, secrets management and OS/container-level sandboxing still need more work.
I'm publishing it because I want those assumptions challenged.
The experiment
The bigger question I'm exploring is:
What should the infrastructure layer underneath autonomous AI agents actually look like?
If you're building coding agents, MCP systems, agent swarms or AI developer tools, I'd love to hear what you think.
GitHub:
Top comments (1)
A multi-agent control layer needs a shared contract more than a shared chat. Each handoff should expose the task, allowed tools, evidence produced, decision owner, and completion test. Without that, the team can look busy while responsibility disappears between planner, implementer, reviewer, and tester.