Connect one agent, watch what your rules would decide, then let them decide. About fifteen minutes.
Create a free workspace. You are its owner.
Under Agents, add your agent and list the tools it may use. Under API keys, create a key for that agent. The key is shown once: store it as a secret, for example in ABB_KEY. A key bound to an agent always acts as that agent.
A work unit is one task your agent is handling, with its own record. Create one under Work units → New and copy its ID from the address bar.
New workspaces start in Watch only mode (see Settings → Protection). Your rules decide and record every action your agent sends, but nothing is blocked. Built-in rules already block deletions, and ask a person before money moves, before data is shared or a message goes outside your organization, and before critical-risk actions.
Before the agent does something, it sends the action to AgentBlackBox:
curl -s -X POST https://agentblackbox.si/api/intercept \
-H "x-agentblackbox-key: $ABB_KEY" \
-H "content-type: application/json" \
-d '{"workUnitId":"<work unit id>","tool":"payments.send",
"actionCategory":"MONEY","riskLevel":"HIGH","amount":4800}'The answer names the decision, for example { "decision": "REQUIRE_APPROVAL", "mode": "shadow", ... }.
ALLOW: go ahead as asked.MODIFY or REDACT: go ahead only with transformedAction from the answer.REQUIRE_APPROVAL: in enforce mode the answer carries approvalId. Wait until a person decides: poll GET /api/approvals/<approvalId> until its status is APPROVED or DENIED.DENY: do not do it. In enforce mode this answer has HTTP status 403.curl -s -X POST https://agentblackbox.si/api/events \
-H "x-agentblackbox-key: $ABB_KEY" \
-H "content-type: application/json" \
-d '{"workUnitId":"<work unit id>","type":"ACTION_EXECUTED","title":"Paid invoice INV-2207"}'The event joins the work unit's tamper-evident record. Open the work unit to see its timeline and audit report.
When the watch-only decisions look right, choose Enforce under Settings → Protection. From then on, blocked actions are refused and approvals are required.
Next: the API reference (rate limits and errors), and integrations to let agents act in Slack, Gmail, Drive, GitHub or Notion through AgentBlackBox.