Inside the ACP Console: My Own Agents, Real Data, Every Screen
A feature list doesn’t show what a control plane looks like once your agents run through it. So here is mine: my own workspace, the agents behind Calafia, running in production. The numbers below come straight from the dashboard as of early July 2026. Nothing is staged.
Four screens: see every call, see what it cost, and set a rule on it.
1. Every agent, every call

This is the first screen I open. Over the last 7 days my workspace shows 19,662 tool calls across 28 agents, $784.63 of model spend, 140 calls denied, 165 errors and 121 PII redactions, each with a small trend line. Below that, every run is one dot, plotted by cost and by runtime. The extreme outliers sit in a gutter so the normal range stays readable.
Before this, “what did all my agents do today, and what did it cost” meant grepping logs across services. Now it’s the landing page.
2. The same agent is never the same twice

Click an agent and you see its spread across runs. This is Claude Code over one week. The median run cost $18.59, but runs ranged from under 1¢ to $501.61. The coefficient of variation is 172%, and the console flags it as erratic. Each run does different work, and the loop re-reads a different amount of context each time.
An agent doesn’t have a fixed cost like a function does. If you only watch the average, you miss the $501 run sitting behind an $18 median. The console shows cost, runtime, model calls and tool calls per agent, across the whole spread.
3. Where the money goes
Update, September 2026: the console now opens cost on its own page.

The Cost page answers four questions in order: how much, on what, is it going up, and what should I change. The last answer is a list of levers measured from your own calls. Each lever has the recorded dollars, one concrete change, and a session that shows it. Examples: side-model calls your harness makes on its own, a prompt prefix that keeps getting rebuilt, long sessions whose last third costs more per turn, and recovery turns after failed or denied calls.

This is the screen I built ACP for. Open one run and the bill splits into the loop and the leaf work.
In one run I opened, 99% of the cost was the orchestration loop: the model re-reading its own growing transcript to decide what to do next. The actual work was nearly free. Web search, memory lookup and sending an email cost nothing in that run. The per-step table shows which steps cost the most. The context-growth chart shows the transcript reaching 59.3k tokens by the end. The “What to optimize” panel suggests pruning context between turns and using fewer, better-specced tools so the loop takes fewer turns.
I call this the loop tax. It’s rarely measured because you have to sit in the call path and price every call to see it. How the loop tax works has the longer version.
4. The controls
For each agent, the console lists every tool it declared, captured from the agent’s own requests before anything runs. Each tool is a row with allow, flag or deny one click away. This one is a background agent. I can deny notifications.sendEmail and the call is stopped whatever the model decides. A poisoned web page can’t talk the agent into an exception, because the rule isn’t in the prompt.
Every workspace starts in audit mode. Calls are recorded and nothing is held except by the safety floor, which covers a short list of destructive commands. When you’ve seen enough traffic, you turn enforcement on. You don’t migrate any policy or rewrite any agents first.
The data isn’t locked in the UI
The activity log, the agent rollups and the policy decisions can be queried through the API, so you can pull them into your warehouse, a Grafana board or a weekly Slack digest. The per-step cost breakdown in the X-ray is console-only for now. (API docs)
Try it on your own agents
It’s the same console you just saw, with your data in it. The free tier covers up to 5 initiating agents with unlimited calls and no card:
curl -sf https://agenticcontrolplane.com/install.sh | bash
If you’d rather not do it alone, book 30 minutes and I’ll wire it into your agent with you, live.
— David
Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
irm https://agenticcontrolplane.com/install.ps1 | iex