How to use OpenClaw to negotiate a SAFE
Two OpenClaw agents negotiated a SAFE through Telegram using signed authority limits, an offer audit trail and private human signing links.
juanfiguera.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Set up a SAFE negotiation demonstration using two OpenClaw agents on [separate machines]. Define each side’s signed mandate with [allowed terms], reject offers outside it, log offers and concessions, and send private signing links to [human principals] only after agreement.
Reach people exploring AI agents
See placements →What happened
The author reports agreement in about forty-five seconds and an executed SAFE with human signatures, a negotiation transcript and cryptographic audit trail.
Reported by juanfiguera on BlogHow the agent did it
- Built APOA as an OpenClaw skill for signed agent mandates.
- Set up two agents on separate machines with Telegram and negotiation boundaries.
- Ran alternating offers in a shared group chat with mandate enforcement.
Where a person stepped in
Humans set each side’s boundaries and signed privately after agreement; the author notes that compromised frameworks remain outside the protection.
A person set the goal and reviewed the result.
Juan Figuera built Agentic Power of Attorney, or APOA, to express agent authority in signed, machine-readable mandates. For a SAFE negotiation demonstration involving a friend's startup, two OpenClaw agents ran on separate machines and used Telegram. Each side supplied boundaries, and the agents negotiated in a shared group chat.
The author implemented an alternating-offers protocol based on Rubinstein's bargaining framework. Offers outside an agent's signed mandate were rejected before reaching the other side. Offers, counteroffers and concessions were recorded through sshsign, and the author reports convergence in about forty-five seconds.
Once agreement was reached, both humans received private signing links rather than links in the group conversation. They signed in the browser, producing an executed SAFE with the negotiation transcript and cryptographic audit trail attached.
The author explicitly limits the claim: this was an early demonstration with well-defined numerical parameters. Enforcement happened at the agent framework layer because services did not support APOA natively, and it would not stop a compromised framework. Who it's for: agent infrastructure builders exploring constrained negotiation demonstrations with explicit boundaries and human signatures.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · juanfiguera
Frequently asked questions
Did the agents sign without human involvement?
No. The author says both humans received private signing links after agreement and drew their signatures in the browser.
Where were negotiation limits enforced?
The framework rejected offers outside signed mandates before forwarding them. The author warns that this does not stop a compromised framework and that native service enforcement was unavailable.
More details
Remaining steps
- Logged offers and concessions through sshsign.
- Sent private signing links after agreement and obtained human signatures.
About this account
The agents exchanged offers within signed mandates; framework checks rejected out-of-bounds offers before delivery.
Platform: Blog. Attribution: https://www.juanfiguera.com/notes/system-prompt-and-vibes/. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: juanfiguera on Blog
Source published 7 May 2026
Source checked 14 Oct 2026
Tools and topics
OpenClaw · APOA · Telegram · sshsign · Ed25519
Agent negotiation · Signed mandates · Audit trails
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.