OpenClaw triages three business email accounts
Jarvis Specter tested an OpenClaw email agent for thirty days, keeping all sends behind approval and narrowing its scope after tone, context and CC mistakes.
dev.to
Original prompt
Shared by the source author. See source notes for context.
Never send without my approval
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Jarvis Specter
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Specter reports useful ongoing triage and routine drafts, with approval gates catching tone, context and CC mistakes before sending.
- Human involvement
- The owner reviews and approves every outgoing message and narrowed the categories eligible for drafting after observing failures.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Connect the three approved accounts and set classification and approval rules.
- Let Donna classify messages, flag urgent items and draft routine replies.
- Review recipients, tone and thread context before approving any send.
Jarvis Specter assigned an OpenClaw agent named Donna to three email accounts through IMAP and Microsoft Graph. It checked during business hours, classified messages, flagged urgent items through Telegram and drafted routine replies. Sending always required the owner's approval.
The thirty-day account reports useful triage and faster review of routine drafts. Persistent memory held relationship notes and corrections. The author also describes a multi-day supplier negotiation where Donna missed a nuance from an earlier session, despite having thread history.
Tone and recipients were recurring problems. Drafts for familiar colleagues sounded generic, and unnecessary CCs could have exposed conversations too broadly if sent. The owner says these issues were caught before embarrassing emails went out. He then limited drafting to routine operations and templated information requests, while flagging negotiations, legal matters and sensitive relationships for personal attention.
The result is a narrower ongoing assistant rather than unattended inbox management. Reported accuracy and time improvements are the author's estimates, with no published measurement set. The concrete failure evidence explains why human approval remained useful: a technically plausible draft could still miss social context or include the wrong people.
Prompt
Original prompt
Source: DEV Community · Jarvis Specter
Frequently asked questions
Could Donna send without approval?
The author explicitly required approval for every send.
What happened in the supplier negotiation?
A later-session draft missed a prior nuance, showing that memory and thread history did not preserve every relevant detail.
How was the scope changed?
Routine operations and templated requests remained eligible; negotiations, legal issues and relationship-sensitive messages were flagged for the owner.
More details
Remaining steps
- Narrow the drafting scope after relationship and negotiation failures.
About this account
OpenClaw classifies inbox messages, flags urgent items in Telegram and drafts routine replies.
Firsthand account opened and read. Original operating-instruction excerpt from the published setup, not the complete task prompt. Results and timings are author reports, not independently reproduced.
Credit: Jarvis Specter
Source published
Added to Agents In Use
Source checked
Tools and topics
OpenClaw · IMAP · Microsoft Graph · Telegram · Memory files
Work · Reported deployment
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.