Build
Build with the agents you already use.
Bring the agents you already use, or build your own loop for your project, and run them side by side. Every agent can see the same map of your architecture, parsed from your code, and commits in its own Git worktree. When an agent finishes, Panolayer checks its work for regressions, and nothing merges into your branch until you approve it.
- Panolayer Agent
- Aider
- Claude Code
- Cline CLI
- Codex CLI
- Crush
- Deep Agents Code
- Gemini CLI
- GitHub Copilot CLI
- Goose
- Kilo Code CLI
- Kimi Code
- Letta Code
- mini-SWE-agent
- Mistral Vibe
- Open Interpreter
- OpenClaw
- OpenCode
- Pi Coding Agent
- Plandex
- Qwen Code
The bug isn’t in the files the agent changed. It’s in the ones that depend on them.
Say you ask an agent to switch getPrice() to return dollars instead of cents. It updates the product page and the cart, and the tests pass. But checkout also calls getPrice(), still treats the result as cents, and a $20 order gets charged 20 cents.
01 / Build
All your agents in one place
Most developers now use more than one agent, each in its own terminal with its own idea of your codebase. In Panolayer, they work side by side, and every one can see the same map of your architecture, parsed from your code, so they all start from the same picture of how the system fits together.
Each agent commits in its own Git worktree, so they don’t overwrite each other, and you see every run, its status, and its verification result in one view.
02 / Build
Install in one click, with your own API keys
Choose from over 20 coding agents and counting in the community-created Agent Marketplace, including Aider, Claude Code, Cline CLI, Codex CLI, Gemini CLI, and GitHub Copilot CLI, plus Panolayer Agent. Install the one you use with one click and start it on your project from the same page, without setting anything up in the terminal.
Agents run on your own subscription or API key. You pay your model provider directly, and Panolayer doesn’t sit in the middle of that bill.
Agent Marketplace03 / Build
Nothing lands without you
An agent can’t merge its own work into your branch. When it’s done, Panolayer verifies the result, and then you decide. Agents from the Marketplace run in a hardened container with controlled workspace scope, tools, network, resource limits, and secrets access. Your main tree isn’t mounted inside it, so the agent has no path to your branch. CLI agents already installed on your Mac start inside their worktree, with instructions to request a merge instead of performing one.
The checkpoint is before your local branch
Your local branch
feature/loginYour starting point. You can work on main, too.
Separate agent worktrees
Codex CLIOwn checkout · own editsAgents commit here as often as they need.
Panolayer checks the work
When the agent declares it is done, verification runs.
↶ Checks fail → back to the agent
Your local branch
feature/loginApproved changes join your work here.
- You push
- Remote feature branch
- Pull request & review
- Merge to main
Your main working tree is not mounted in the container. The agent cannot merge directly into it.
Run on your Mac in a worktree. They are instructed to request a merge; this is policy, not container isolation.
You can always commit and merge agent work yourself, even when checks fail.
04 / Build
Every run is checked
An agent’s summary of what it did isn’t always what it did. When an agent finishes, Panolayer records every file it actually changed, not just the ones it mentioned, and checks the whole change against the rules on that code and on everything that depends on it.
How verification works05 / Build
Build your own agent
An agent’s behavior comes from its harness: how it plans, what it checks before it says it’s done, and what it does when a step fails. Usually you can only customize the prompt. The Graph Builder lets you design the harness itself by laying it out as a graph. When you’re done, you can keep it private or publish it to the Marketplace.
How Graph Builder worksPanolayer Agent
Panolayer’s own agent, built into the platform and connected to the architecture map and verification. Use it to code, the same way you’d use any other agent.
When verification flags a regression, Panolayer Agent takes the failure and the affected callers as its brief, fixes it in its own worktree, and the fix goes through verification again.
It’s also what runs the loops you build. A loop is the workflow you lay out in the Graph Builder: the steps, the checks between them, and what happens when one fails. Panolayer Agent works through it step by step, runs each check you added, and follows the failure path you set when something goes wrong.
And when you’re reviewing changes before a commit, you can open a chat with it with the diff already loaded, and ask what changed and why instead of reading every line.
Questions
Which agents are supported?
Panolayer Agent and a growing selection of third-party agents are supported. Browse the Agent Marketplace for the latest additions and the agent reference for setup details.
See all supported agentsDoes every agent run in a container?
No. Agents you install from the Marketplace do. CLI agents already installed on your Mac run as you, in their own worktree. That keeps their work separate from your branch, but it isn’t a security sandbox.
Can an agent merge its own work?
No. Its work is verified first, and then you decide. You can merge it yourself even with checks failing. The failures stay listed in Verify.
Do agents see my architecture?
They can. Panolayer gives agents an MCP tool for querying its map of your codebase: structure, callers, and dependencies. Whether an agent uses it on a given task is up to the agent.
Can I run more than one agent at once?
Yes. Each gets its own worktree, and you see every run side by side.
Do I need my own subscription or API key?
Yes. Agents run on your own subscription or API key. Panolayer doesn’t resell model access.
Does this replace my editor?
No. Panolayer runs beside it. Agents work in their own worktrees, so you can keep working in your editor while they run.