Replay any run
Run a past task again, step by step, against the same inputs.
AI agent control plane
Signalbox routes every AI agent you run in production, records each action in one register, and lets one person hold them all.
Sample names. Yours could be here.
To begin, install the SDK and wrap the agent you already have. No rewrite, no new framework.
On its next task, every run lands in the register with its route, cost and status.
From then on, your rules set the routes. An agent can only take the routes they allow.
Signal boxes have used interlocking since the 1850s. Set one route and the levers for any crossing route lock. Signalbox does the same for AI agents: guardrails that block conflicting actions before they run.
Two agents want order 48213.
refunds-agent v14 sets its route first. It locks.
chargeback-agent v8 gets a red signal. It waits, on the record.
Six things you can change while the agents are running.
Run a past task again, step by step, against the same inputs.
Set a spend limit per agent. At the cap, its runs wait at a signal.
Put a person on any step, and the run waits until they say yes.
Change the model behind an agent in config, and the next run uses it.
Names, emails and card numbers are masked before a prompt reaches the model.
Get a message in the right channel when a run waits, fails or hits a cap.
Mark the steps that need a person. The run stops there, and moves when someone says so.
The run reaches a step you marked for review.
It stops. Nothing moves until someone approves.
Approved. The run finishes, and the register keeps who said yes.
A refund over the £100 limit waits for a support lead.
A matched invoice waits for a controller before payment is booked.
Admin rights on a billing workspace wait for the system owner.
A discount past the 15% rule waits for the deal desk.
The register is your audit log, and the record. Here is how it is kept, and who can touch it.
Your whole team can watch the board. You pay for the runs it records.
One team putting its first agents into production.
Several teams, several agents, rules that cross between them.
Agents across regions, with self-hosting and audit needs.
Every plan: unlimited seats, the hold lever and Slack alerts. Prices and limits are placeholders on this concept site.
Signalbox is a control plane for AI agents in production. It routes each agent task, records every action in a signed register, blocks conflicting actions with interlocking rules, and lets one person hold any run, or all of them, with a single lever.
Observability tools show you what an agent did. Signalbox also decides what it may do next. Each action is checked against your rules before it runs, so a conflict is stopped at the signal, not found in a dashboard the next morning.
Signalbox wraps the agent, not the model, so it works with the stack you already run. Changing the model behind an agent is a config change. The next run uses it, with no redeploy.
Every live run stops before its next step and waits. Nothing is cancelled and no state is lost. The register logs who held the runs and when. Pull the lever again and each run carries on from the step where it stopped.
Yes, on the Network plan. Signalbox runs inside your own VPC, so prompts, outputs and the register stay on your network. On Depot and Junction we host it, with your register pinned to the region you choose.
A call with an engineer. We connect a sample agent, put its runs on the board, set one interlocking rule and pull the lever. Bring the workflow that worries you most, and we’ll map its routes with you.
Signal cleared. We’ll email you to find a time.
That didn’t go through. Check the address and try again.
We’ll email you once to find a time. This is a concept site, so please don’t send real details.