Decide whether an agent's next action needs approval
Before an AI agent sends an email, changes a record, or spends money, a decision model rates how risky the action is, lets low-risk actions run, and pauses everything else for a person to approve.
Author's figure: Author's figure: about $0.04 a day to check every Claude Code tool call (Bouncer). Reported by the builder, not measured by us.
The use case
Agents that draft and send emails, update CRM records, or create invoices are useful until one does something hard to undo. Jev reads the proposed action, its target, and the content, and answers how risky the action is and whether it is reversible. Actions rated low risk and reversible with a confidence of 0.9 or higher run on their own, such as adding an internal note. Anything else is paused and sent to a person in Slack with an approve or reject button. Payments, refunds, and messages to customers always require approval no matter what Jev says. A person decides every paused action, and the log shows what ran and why.
Example workflow
- The agent proposes an action as a small JSON object with the action type, target, and content.
- Before running it, the agent's code sends that object to Jev as the state.
- Jev answers a Score question, how risky is this action, with the levels low, medium, and high, and a Noul question, this action can be easily undone.
- Code lets the action run only when risk is low, reversible is likely, confidence clears the threshold, and the action type is not on the always-approve list.
- Every other action is posted to Slack with approve and reject buttons, and the agent waits.
- Every decision, Jev answer, and approval is logged in a Google Sheet for weekly review.
Tools
- Jev (TypeSafe AI)
- Claude
- Slack
- Zapier or Make
- Google Sheets
- HubSpot
Risks and guardrails
- A risky action rated low could run without review — guardrail: payments, refunds, deletions, and customer messages are on a fixed always-approve list in code.
- Content injected into the agent could try to talk the check into a pass — guardrail: the action type is checked in code, not by Jev, and Jev only sees the proposed action.
- Too many pauses make the agent useless — guardrail: review paused actions weekly and move clearly safe types to the allowed list by hand.
A similar pattern in a public build: https://madewithjev.com/builds/agent-safety-monitor. See also Decision models.
Related recipe
B2B Invoice AR Reminder Agent with Approval Gate: An agent reads open invoices from QuickBooks or Xero, skips blocked accounts, and drafts tiered payment reminders in Gmail that a person approves one by one before anything is sent.
Want it set up for you? Request a quote.
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Public build with a similar pattern: madewithjev.com/builds/agent-safety-monitor
- Bouncer by Clownware on Made with Jev: madewithjev.com/builds/bouncer
- Bouncer: original post or repo by Clownware: github.com/clownware/bouncer
- Jev as an agent safety monitor: original post or repo by Nick: x.com/isNickMa/status/2100566407524344225
- Agent Chaperone by agent-chaperone on Made with Jev: madewithjev.com/builds/agent-chaperone
- Agent Chaperone: original post or repo by agent-chaperone: github.com/agent-chaperone/agent-chaperone
- An agent-agnostic permission gate by ImRobot on Made with Jev: madewithjev.com/builds/agent-permission-gate
- An agent-agnostic permission gate: original post or repo by ImRobot: x.com/th3nolo/status/2103024142165357016
- jev-permission-gate by Madison Rickert on Made with Jev: madewithjev.com/builds/jev-permission-gate
- jev-permission-gate: original post or repo by Madison Rickert: github.com/madisonrickert/jev-permission-gate
- Jev blocks a $50,000 transfer by fluixoo on Made with Jev: madewithjev.com/builds/jev-blocks-transfer
- Jev blocks a $50,000 transfer: original post or repo by fluixoo: x.com/fluixoo/status/2102680895425503354
- pi-heed by Bowen Xu on Made with Jev: madewithjev.com/builds/pi-heed
- pi-heed: original post or repo by Bowen Xu: github.com/Nyarlathoteppppp/pi-heed
- Interlock by somoore on Made with Jev: madewithjev.com/builds/interlock
- Interlock: original post or repo by somoore: github.com/somoore/interlock
- A Jev gate belongs in a hook, not MCP by Automater on Made with Jev: madewithjev.com/builds/jev-claude-code-hook
- A Jev gate belongs in a hook, not MCP: original post or repo by Automater: automater.ai/intel/jev-claude-code-hook-not-mcp
- Every rule checked in 348 ms by erkamyaman_ng on Made with Jev: madewithjev.com/builds/rule-enforcement-claude-code
- Every rule checked in 348 ms: original post or repo by erkamyaman_ng: x.com/erkamyaman_ng/status/2102759754548314325
- Show HN post by the agent-chaperone author: news.ycombinator.com/item
- Hacker News comment by gbrindisi on moving prompt-injection checks to Jev: news.ycombinator.com/item
- Claude (official page): claude.ai (checked 2026-11-09)
- Slack MCP server (official page): docs.slack.dev/ai/slack-mcp-server (checked 2026-11-09)
- Zapier MCP (official page): zapier.com/mcp (checked 2026-11-09)
- Make MCP server (official page): developers.make.com/mcp-server (checked 2026-11-09)
- HubSpot MCP server (official page): developers.hubspot.com/ai-tools/mcp (checked 2026-11-09)